APASA: adaptive selection of informative peritumoral regions for improved automated cancer lesion analysis Authors: Xiaoyang Duan, Patrice Monkam, Xu Wang, Yabin Zhang, Shouliang Qi, Dan Zhao, Tao Yu, Mkiramweni Mbazingwa Elirehema, Deogratias Mzurikwao, Wei Qian Published: 2026 ABSTRACT Precise cancer lesion analysis in medical imaging critically depends on the accurate definition of regions of interest (ROIs), which directly influence diagnostic and clinical outcomes. While peritumoral features are known to enhance lesion characterization, efficiently defining meaningful peritumoral ROIs remains a challenge. We propose an adaptive peritumoral area selection approach (APASA) that systematically identifies the most informative ROI surrounding a lesion, enabling the extraction of meaningful radiomic features for improved diagnostic performance. Unlike conventional heuristic or morphology-based methods, APASA leverages the minimum coverage graph algorithm, using the tumor ROI as a reference to construct a graph encompassing both the tumor and its peritumoral microenvironment. The effectiveness of the proposed approach was evaluated within AI-based frameworks for automated lesion differentiation in breast and thyroid cancers. Extensive experiments employing five widely used machine learning models demonstrated that APASA-selected peritumoral features consistently outperformed conventional morphological dilation. Performance improvements reached up to 30.75% in AUC and 29.00% in F1-score compared with the tumor ROI baseline. Moreover, the optimal model was found to vary depending on the ROI type, shape, and cancer type, offering new insights into the interaction between ROI selection and model choice. These results highlight APASA as a principled and efficient strategy for adaptive ROI definition in ultrasound-based cancer lesion analysis, demonstrating effectiveness across two ultrasound datasets, with potential extension to other imaging modalities and clinical settings pending further validation. Introduction Cancer constitutes a major global public health challenge and continues to be one of the leading causes of mortality worldwide1,2. According to the latest global estimates from the World Health Organization (WHO), cancer is responsible for approximately 9 to 10 million deaths annually, accounting for nearly one in six deaths globally3. Depending on the tissue or organ affected by malignant cells, cancer can be categorized into various types, including breast cancer, thyroid cancer, lung cancer, prostate cancer, liver cancer, etc. Among these, breast and thyroid cancers are of particular interest due to their significant health impact and epidemiological characteristics. Breast cancer remains the most commonly diagnosed malignant tumor among women. According to recent global estimates, approximately 2.3 million new female breast cancer cases and 670,000 related deaths occurred worldwide in 2022, and the annual burden is projected to rise to 3.2 million new cases and 1.1 million deaths by 2050 if current rates continue4,5. Thyroid cancer is also a common endocrine malignancy with an increasing global burden. Based on GLOBOCAN 2022 estimates, approximately 821,214 new thyroid cancer cases and 47,507 thyroid cancer-related deaths occurred worldwide in 2022, with about 64.63% of cases occurring in individuals younger than 55 years6. Therefore, increasing efforts have been devoted to achieving early and accurate diagnosis and developing personalized treatment strategies, both of which are essential for improving patient outcomes and reducing mortality rates. Early cancer diagnosis and treatment planning are typically conducted through routine screening procedures that rely on advanced technologies, with medical imaging being the most commonly adopted category. A variety of imaging modalities are available, including X-ray, ultrasound (US), computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET)7,8. Among these, ultrasound stands out as a widely used technique due to its non-invasive nature, lack of ionizing radiation, convenience, and cost-effectiveness; attributes that make it particularly suitable for screening and diagnosing breast and thyroid cancers8,9. However, the effective analysis of imaging data remains hindered by several challenges, including the limited availability of trained physicians and the poor reproducibility resulting from inter- and intra-observer variability10. For instance, in current clinical practice, it typically requires 5 to 10 years of specialized training at substantial cost to cultivate an experienced sonographer. Meanwhile, the average annual growth rate of ultrasound examinations has exceeded 15%, driven by the increasing demand for precision medicine11. To address challenges related to cancer data analysis, typically including inter- and intra-observer variability and imbalance between available physicians and the growing volume of data, numerous automated tools have been developed12–14. Among these, approaches based on artificial intelligence (AI) have emerged as particularly promising, demonstrating strong potential to enhance diagnostic accuracy, efficiency, and consistency15–17. One key task in this data analysis procedure is lesion differentiation, which constitutes a foundational step in cancer detection and analysis. In the context of AI-driven approaches for automated lesion differentiation, models are typically developed based on the lesion area, often obtained through steps such as segmentation or region of interest (ROI) detection. However, given the clinical importance of peritumoral microenvironment features, numerous researchers have begun to explore the value of incorporating such contextual information to enhance the differentiation performance of AI-based methods. In recent years, numerous studies have investigated the integration of peritumoral features to enhance the performance of AI-based cancer analysis. These works have explored different strategies to incorporate or compare peritumoral and tumoral information across various clinical tasks, such as lesion classification5,18–22 and treatment response prediction23–26. A range of imaging modalities, including ultrasound, MRI, CT, and mammography, have been employed, demonstrating the versatility and clinical relevance of peritumoral features. Several researchers have proposed approaches that incorporate peritumoral information to differentiate between benign and malignant breast lesions across various imaging modalities, including ultrasound (US)5,18, mammography19,20, and multimodal data21,22. Other studies have focused on leveraging peritumoral features to improve the precision and efficacy of treatment response prediction, using US data23,24 and MRI data25,26. Beyond breast cancer, the clinical value of peritumoral features has also been explored in other cancer types. For example, lesion differentiation in brain cancer using MRI data27, genetic mutation prediction in lung lesions using CT data28, treatment response prediction in liver cancer using both MRI and CT data29–31, and in prostate cancer using MRI data32. With the continuous advancement of frontier technologies, computer-aided diagnosis (CAD) systems have progressively evolved from traditional machine learning-based approaches relying on handcrafted features to AI-driven frameworks for automated lesion analysis. These systems have been widely applied in clinical practice for tasks such as lesion detection, classification, and treatment response assessment, thereby improving diagnostic efficiency. However, their diagnostic performance remains limited by the lack of an efficient and adaptive strategy for defining peritumoral regions, as current approaches often rely on non-standardized morphological dilation operations that are sensitive to lesion morphology. Therefore, developing an adaptive peritumoral region selection method would be highly beneficial for improving the diagnostic performance, stability, and clinical applicability of CAD systems. For the reader’s convenience, we have summarized representative studies involving tumors and their surrounding regions in Table 1.Table 1Summary of representative studies using tumoral and peritumoral information for cancer lesion analysis.StudyCancer type and clinical taskImaging modalityROI usage methodCui et al.18Breast cancer; benign/malignant classificationUltrasoundFused intratumoral and peritumoral regions in a multi-region network.Guo et al.19Breast cancer; lesion classificationMammographyCompared tumoral, peritumoral, and combined ROIs for radiomics/deep learning analysis.Zhang et al.20Breast cancer; malignant BI-RADS category 4 lesion predictionMammographyUsed intra- and peritumoral radiomics features from predefined surrounding regions.Wang et al.21Breast cancer; HER2 status predictionUltrasoundExtracted radiomics features from tumor and 5-mm peritumoral ROIs.Tian et al.22Breast cancer; benign/malignant classificationDBT-CC and UltrasoundIntegrated lesion-centered multimodal features for classification model construction.Wu et al.23Breast cancer; Ki-67 expression predictionUltrasoundBuilt a nomogram using intra- and 10-mm peritumoral radiomics features.Wang et al.24Breast cancer; Ki-67 predictionUltrasoundDerived peritumoral ROI by 3-mm radial expansion of the tumor boundary.Zheng et al.25,26Breast cancer; pathological complete response predictionMRIUsed multi-region MRI radiomics from intratumoral and tumor-associated surrounding tissues.Yan et al.27Brain cancer; glioblastoma and solitary brain metastasis differentiationMRI and DWIUsed lesion-centered MRI/DWI regions for deep learning-based differentiation.Omura et al.28Lung cancer; EGFR mutation predictionCTExtracted features from tumor, tumor rim, and tumor exterior regions.Liu et al.29Liver cancer; pathological differentiation predictionMRICompared intratumoral and multiregional peritumoral radiomics models.Wei et al.30Liver cancer; microvascular invasion predictionMRICompared tumor-region features with 5-, 10-, and 20-mm peritumoral features.Shan et al.31Liver cancer; early recurrence predictionCTUsed tumoral ROI and an additional 2-cm peritumoral ROI.Algohary et al.32Prostate cancer; treatment response evaluationMRIAnalyzed longitudinal regional MRI features for treatment response assessment.Xu et al.33Breast cancer; lesion segmentation and classificationUltrasoundUsed regional attention to learn tumor, peritumoral, and background information. Table 1Summary of representative studies using tumoral and peritumoral information for cancer lesion analysis.StudyCancer type and clinical taskImaging modalityROI usage methodCui et al.18Breast cancer; benign/malignant classificationUltrasoundFused intratumoral and peritumoral regions in a multi-region network.Guo et al.19Breast cancer; lesion classificationMammographyCompared tumoral, peritumoral, and combined ROIs for radiomics/deep learning analysis.Zhang et al.20Breast cancer; malignant BI-RADS category 4 lesion predictionMammographyUsed intra- and peritumoral radiomics features from predefined surrounding regions.Wang et al.21Breast cancer; HER2 status predictionUltrasoundExtracted radiomics features from tumor and 5-mm peritumoral ROIs.Tian et al.22Breast cancer; benign/malignant classificationDBT-CC and UltrasoundIntegrated lesion-centered multimodal features for classification model construction.Wu et al.23Breast cancer; Ki-67 expression predictionUltrasoundBuilt a nomogram using intra- and 10-mm peritumoral radiomics features.Wang et al.24Breast cancer; Ki-67 predictionUltrasoundDerived peritumoral ROI by 3-mm radial expansion of the tumor boundary.Zheng et al.25,26Breast cancer; pathological complete response predictionMRIUsed multi-region MRI radiomics from intratumoral and tumor-associated surrounding tissues.Yan et al.27Brain cancer; glioblastoma and solitary brain metastasis differentiationMRI and DWIUsed lesion-centered MRI/DWI regions for deep learning-based differentiation.Omura et al.28Lung cancer; EGFR mutation predictionCTExtracted features from tumor, tumor rim, and tumor exterior regions.Liu et al.29Liver cancer; pathological differentiation predictionMRICompared intratumoral and multiregional peritumoral radiomics models.Wei et al.30Liver cancer; microvascular invasion predictionMRICompared tumor-region features with 5-, 10-, and 20-mm peritumoral features.Shan et al.31Liver cancer; early recurrence predictionCTUsed tumoral ROI and an additional 2-cm peritumoral ROI.Algohary et al.32Prostate cancer; treatment response evaluationMRIAnalyzed longitudinal regional MRI features for treatment response assessment.Xu et al.33Breast cancer; lesion segmentation and classificationUltrasoundUsed regional attention to learn tumor, peritumoral, and background information. Summary of representative studies using tumoral and peritumoral information for cancer lesion analysis. Numerous studies have demonstrated the clinical value of peritumoral microenvironment features in enhancing the precision and effectiveness of various tasks across different cancer types. These efforts have highlighted the significant contributions of peritumoral information to advancing AI-driven precision oncology. Despite these advancements, several critical challenges remain. First, the peritumoral region is often defined using morphological operations, which may inadvertently alter key tumor shape features that are crucial for clinical decision-making. Second, many studies rely on labor-intensive empirical analyses to determine the optimal expansion range for capturing informative peritumoral features, underscoring the urgent need for a standardized approach to peritumoral region selection. These limitations not only hinder reproducibility but also constrain the clinical translation and scalability of such AI-based approaches. Therefore, developing a strategy that enables streamlined and standardized selection of peritumoral information, without relying on morphological operations, would have important clinical implications. That is, the development of an adaptive, reproducible, and morphology-independent ROI selection strategy for peritumoral feature extraction is essential for improving AI-based ultrasound breast lesion analysis and facilitating more reliable automated diagnostic decision-making. To address limitations in existing peritumoral feature extraction methods, this study explores the adaptive selection of peritumoral microenvironment information without relying on morphological operations or labor-intensive empirical definitions. Specifically, we introduce a novel method, termed APASA (adaptive peritumoral area selection approach), designed to identify informative peritumoral regions in a standardized and efficient manner. Inspired by the minimum coverage graph algorithm, a graph-based method for defining minimal enclosing regions, APASA uses the intratumoral (or tumor) region as a reference to construct a minimum coverage graph that encompasses both the tumor and its surrounding peritumoral area. This formulation enables an adaptive and reproducible definition of peritumoral regions for enhanced lesion analysis. The effectiveness of the proposed APASA strategy is evaluated on lesion differentiation tasks in two cancer types: breast and thyroid cancers. Radiomics features extracted from the regions defined using APASA are used to train five commonly used machine learning models, and the resulting classification performances are compared against those achieved using conventional peritumoral selection methods. Experimental results demonstrate that our adaptive strategy consistently outperforms existing approaches, highlighting its robustness and effectiveness. We believe that APASA opens new avenues for automated lesion analysis by offering a standardized and reliable framework to enhance early cancer diagnosis and treatment planning. The key contributions of our work can be summarized as follows:Introduction of a novel adaptive strategy for appropriate peritumoral region selection. We introduce APASA, an innovative method inspired by the minimum coverage graph algorithm, which enables standardized and morphology-independent selection of informative peritumoral regions for cancer lesion analysis.Improved lesion differentiation through adaptive peritumoral feature extraction. By applying our proposed APASA strategy to breast and thyroid cancer datasets, it enables more effective radiomics-based feature extraction, achieving superior classification performance compared to conventional peritumoral selection methods.Insights into the interplay between ROI selection and model performance. Experiments reveal that the optimal ML model varies depending on ROI type, shape, and cancer type, highlighting the importance of tailoring model selection to specific ROI configurations and clinical scenarios.Development of a clinically relevant AI-driven framework for automated cancer analysis. Extensive experimental analyses demonstrate that integrating the proposed strategy offers a potentially scalable approach for automated lesion analysis in breast and thyroid cancers, and suggests its potential applicability to broader lesion analysis settings. Introduction of a novel adaptive strategy for appropriate peritumoral region selection. We introduce APASA, an innovative method inspired by the minimum coverage graph algorithm, which enables standardized and morphology-independent selection of informative peritumoral regions for cancer lesion analysis. Introduction of a novel adaptive strategy for appropriate peritumoral region selection. We introduce APASA, an innovative method inspired by the minimum coverage graph algorithm, which enables standardized and morphology-independent selection of informative peritumoral regions for cancer lesion analysis. Improved lesion differentiation through adaptive peritumoral feature extraction. By applying our proposed APASA strategy to breast and thyroid cancer datasets, it enables more effective radiomics-based feature extraction, achieving superior classification performance compared to conventional peritumoral selection methods. Improved lesion differentiation through adaptive peritumoral feature extraction. By applying our proposed APASA strategy to breast and thyroid cancer datasets, it enables more effective radiomics-based feature extraction, achieving superior classification performance compared to conventional peritumoral selection methods. Insights into the interplay between ROI selection and model performance. Experiments reveal that the optimal ML model varies depending on ROI type, shape, and cancer type, highlighting the importance of tailoring model selection to specific ROI configurations and clinical scenarios. Insights into the interplay between ROI selection and model performance. Experiments reveal that the optimal ML model varies depending on ROI type, shape, and cancer type, highlighting the importance of tailoring model selection to specific ROI configurations and clinical scenarios. Development of a clinically relevant AI-driven framework for automated cancer analysis. Extensive experimental analyses demonstrate that integrating the proposed strategy offers a potentially scalable approach for automated lesion analysis in breast and thyroid cancers, and suggests its potential applicability to broader lesion analysis settings. Development of a clinically relevant AI-driven framework for automated cancer analysis. Extensive experimental analyses demonstrate that integrating the proposed strategy offers a potentially scalable approach for automated lesion analysis in breast and thyroid cancers, and suggests its potential applicability to broader lesion analysis settings. Methodology Overall pipeline of the proposed framework for automated cancer lesion analysis In this study, we leverage the proposed adaptive peritumoral area selection approach (APASA) to develop a clinically relevant AI-driven framework for automated cancer lesion analysis. The overall development pipeline, illustrated in Fig. 1, comprises several key stages: which involves identification of tumoral regions and adaptive selection of their corresponding peritumoral areas using APASA; radiomics feature extraction and selection; and ML-based classification and evaluation. This integrated process enables standardized, morphology-independent utilization of peritumoral information, ultimately leading to more accurate and reproducible lesion differentiation. Detailed descriptions of each stage in the pipeline are provided in the following subsections.Fig. 1Overall pipeline of the developed AI-based framework for automated cancer lesion differentiation. Overall pipeline of the developed AI-based framework for automated cancer lesion differentiation. Data preparation: tumor region identification and APASA-based peritumoral selection Data preparation plays a crucial role in the successful implementation and reliability of AI-based automated frameworks. In cancer lesion analysis, it generally involves two main steps: (1) tumor region identification, which defines both the intratumoral (tumor) and peritumoral regions, and (2) lesion type classification, determining whether the lesion is benign or malignant. The effectiveness of our proposed framework is validated on two cancer types: breast and thyroid cancers. The thyroid cancer dataset was obtained from a public source, with all lesion type labels and tumor region delineations already provided. For the breast cancer dataset, preparation was carried out by two experienced radiologists. After removing device-specific and patient-identifiable information from the ultrasound images, puncture biopsy examinations were used to determine each lesion’s type (benign or malignant). Next, the image showing the most informative lesion details was selected for each patient, and tumor regions were manually delineated by each radiologist. The delineations were then cross-checked to ensure accuracy and consistency. Once both datasets were ready, including images, tumor masks, and lesion type labels, we applied specific operations to identify multiple regions of interest (ROIs). For each image and its tumor mask, our proposed APASA was applied to construct a minimum coverage graph (MCG). The MCG adaptively encompassed both the tumor and its surrounding peritumoral area. This process required no morphological operations and relied solely on the tumor’s boundary coordinates. From the MCG, the peritumoral region was obtained by subtracting the tumor mask. As a result, three masks were generated for each case: (1) the tumor mask, (2) the MCG mask (tumor + peritumoral area), and (3) the peritumoral mask. Pixel-wise multiplication between each mask and the corresponding image produced three distinct ROIs, which were subsequently used for radiomics feature extraction. It is important to note that the construction of the minimum coverage graph in our proposed APASA framework allows for flexibility in the choice of shape. In this study, we investigate two shape configurations: rectangular and circular. These two shapes were selected to investigate the influence of peritumoral region geometry on the extracted radiomics features and subsequent classification performance, thereby assessing the robustness and adaptability of the proposed method across different spatial representations. The overall ROI generation process is illustrated in Fig. 2.Fig. 2Workflow diagram for the overall ROI generation process. Workflow diagram for the overall ROI generation process. Overview and implementation of our proposed APASA Accurate and reproducible delineation of both intratumoral and peritumoral regions is critical for reliable cancer lesion differentiation. Conventional peritumoral selection methods, such as fixed-distance expansions or morphological dilation, often depend on the lesion’s shape, making them less robust for tumors with irregular boundaries. In particular, previous studies using morphological expansion have faced challenges in automatically and precisely determining the optimal peritumoral range in a timely manner. To overcome these limitations, we introduce APASA, an innovative strategy that automatically determines optimal peritumoral regions through the construction of a minimum coverage graph (MCG). This morphology-independent method standardizes ROI generation across diverse lesion types and shapes, thereby enhancing reproducibility and clinical applicability. The core of our method is based on the principles of computational geometry. Given a set of points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{p_1,p_2,...,p_n\}$$\end{document} within a defined planar region, the objective is to determine a geometric shape G that satisfies the following conditions: Complete coverage: All points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i \in P$$\end{document} lie inside or on the boundary of the geometry G.Minimal coverage: Among all geometric shapes satisfying the complete coverage condition, G is the one that encloses all points in P while minimizing the total coverage area.When constructing the MCG, different graph types can be employed, and the implementation details may vary accordingly. In this study, we focus on two specific types: the minimum coverage rectangle (MCR) and the minimum coverage circle (MCC), which serve as representative examples to illustrate the underlying construction principles. The implementation principle for both MCR and MCC are displayed in Fig. 3.Fig. 3Schematic illustration of the minimum coverage graph (MCG) algorithm. (a) Schematic diagram of MCR selection for point set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{a, b, c, d\}$$\end{document}; (b) Schematic diagram of MCC selection for point set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{p_1, p_2, ..., p_n\}$$\end{document}. Complete coverage: All points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i \in P$$\end{document} lie inside or on the boundary of the geometry G. Complete coverage: All points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i \in P$$\end{document} lie inside or on the boundary of the geometry G. Minimal coverage: Among all geometric shapes satisfying the complete coverage condition, G is the one that encloses all points in P while minimizing the total coverage area. Minimal coverage: Among all geometric shapes satisfying the complete coverage condition, G is the one that encloses all points in P while minimizing the total coverage area. Schematic illustration of the minimum coverage graph (MCG) algorithm. (a) Schematic diagram of MCR selection for point set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{a, b, c, d\}$$\end{document}; (b) Schematic diagram of MCC selection for point set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{p_1, p_2, ..., p_n\}$$\end{document}. Principle of peritumoral ROI selection using MCR For the rectangular shape (MCR), consider a set of points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{a, b, c, d\}$$\end{document} with known coordinates. The goal is to find the smallest axis-aligned rectangle that completely encloses all points in P. This is achieved by identifying the coordinate extremes through traversal of all points as follows:The minimum horizontal coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}$$\end{document}) corresponds to point b.The maximum horizontal coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}$$\end{document}) corresponds to point c.The minimum vertical coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{min}$$\end{document}) corresponds to point d.The maximum vertical coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{max}$$\end{document}) corresponds to point a.Once these extremes are determined, the vertices of the MCR are defined as follows: The coordinates of vertex A in the upper left corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}, y_{max}$$\end{document}).The coordinates of vertex B in the upper right corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}, y_{max}$$\end{document}).The coordinates of vertex C in the lower right corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}, y_{min}$$\end{document}).The coordinates of vertex D in the lower left corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}, y_{min}$$\end{document}). The minimum horizontal coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}$$\end{document}) corresponds to point b. The minimum horizontal coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}$$\end{document}) corresponds to point b. The maximum horizontal coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}$$\end{document}) corresponds to point c. The maximum horizontal coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}$$\end{document}) corresponds to point c. The minimum vertical coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{min}$$\end{document}) corresponds to point d. The minimum vertical coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{min}$$\end{document}) corresponds to point d. The maximum vertical coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{max}$$\end{document}) corresponds to point a. The maximum vertical coordinate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{max}$$\end{document}) corresponds to point a. The coordinates of vertex A in the upper left corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}, y_{max}$$\end{document}). The coordinates of vertex A in the upper left corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}, y_{max}$$\end{document}). The coordinates of vertex B in the upper right corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}, y_{max}$$\end{document}). The coordinates of vertex B in the upper right corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}, y_{max}$$\end{document}). The coordinates of vertex C in the lower right corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}, y_{min}$$\end{document}). The coordinates of vertex C in the lower right corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{max}, y_{min}$$\end{document}). The coordinates of vertex D in the lower left corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}, y_{min}$$\end{document}). The coordinates of vertex D in the lower left corner are (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{min}, y_{min}$$\end{document}). Principle of peritumoral ROI selection using MCC For the circular shape (MCC), consider a set of points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{p_1, p_2, ..., p_n\}$$\end{document}. The goal is to determine the smallest possible circle that completely encloses all points in P. This problem, first introduced by James Joseph Sylvester in 185734, has a unique solution for any given finite point set. Notably, at most three points from P lie on the boundary of the MCC. If three points define the boundary, the MCC is their circumcircle. If only two points define the boundary, the MCC is the circle whose diameter is the segment connecting them. A widely used solution is Welzl’s algorithm35, a randomized incremental method based on linear programming principles: Maintain two sets: P (points yet to be processed) and R (points that define the current boundary).Start with an empty boundary circle and process points from P in random order.If a new point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} lies inside or on the current circle, continue without change.If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} lies outside, update the MCC so that it passes through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and one or two points from R: With one point from R, the MCC is the circle with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and that point as endpoints of a diameter; with two points from R, the MCC is the circumcircle defined by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and the two points.Repeat until all points in P have been processed; the resulting circle is the MCC enclosing the entire set. Maintain two sets: P (points yet to be processed) and R (points that define the current boundary). Maintain two sets: P (points yet to be processed) and R (points that define the current boundary). Start with an empty boundary circle and process points from P in random order. Start with an empty boundary circle and process points from P in random order. If a new point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} lies inside or on the current circle, continue without change. If a new point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} lies inside or on the current circle, continue without change. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} lies outside, update the MCC so that it passes through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and one or two points from R: With one point from R, the MCC is the circle with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and that point as endpoints of a diameter; with two points from R, the MCC is the circumcircle defined by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and the two points. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} lies outside, update the MCC so that it passes through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and one or two points from R: With one point from R, the MCC is the circle with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and that point as endpoints of a diameter; with two points from R, the MCC is the circumcircle defined by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} and the two points. Repeat until all points in P have been processed; the resulting circle is the MCC enclosing the entire set. Repeat until all points in P have been processed; the resulting circle is the MCC enclosing the entire set. Application of the MCG algorithm in the analysis of cancer lesions Building on the above principles of the MCG algorithm, we now illustrate its adaptation to the case of cancer lesion analysis. In clinical practice, radiologists typically define key points to determine the location and size of tumors during assessment. After locating the tumor within the image, a tumor mask is generated, and the MCG algorithm is applied to obtain the corresponding MCG mask. A representative example is shown in Fig. 4 for both MCR and MCC, illustrating their application to a delineated tumor.Fig. 4APASA-based tumoral region selection using the MCG algorithm. (a) Schematic of MCR-based region selection. (b) Schematic of MCC-based region selection. APASA-based tumoral region selection using the MCG algorithm. (a) Schematic of MCR-based region selection. (b) Schematic of MCC-based region selection. Radiomics feature extraction and integration Leveraging our proposed APASA strategy and the initially generated tumor mask, three distinct mask types are derived: the tumor mask, the MCG mask (covering the tumor and peritumoral area), and the peritumoral mask. For each mask, the corresponding ROI is obtained by performing pixel-wise multiplication between the mask and its associated image. Prior to this step, all images were normalized, with pixel spacing resampled to [1,1]. Subsequently, the feature extractor function of the PyRadiomics library was applied to each ROI, yielding a total of 1,032 conventional radiomics features. These include shape-based (2D) features, first-order statistical features, and texture features. The texture features consist of gray level co-occurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and neighboring gray tone difference matrix (NGTDM). In addition to these, PyRadiomics computes the corresponding features on multiple transformed images, including exponential, gradient, local binary pattern (2D LBP), logarithmic, square, square root, and wavelet transformations, among others. A comprehensive description of all extracted features can be found in the PyRadiomics documentation (https://pyradiomics.readthedocs.io/en/latest/features.html). To eliminate redundancy and remove less informative features, a two-stage feature selection process was employed. First, an independent samples t-test was used for preliminary selection. This univariate statistical method assesses whether a feature can discriminate between classes by comparing the difference in means across categories. Next, the features retained from the t-test were used as the initial candidate set for least absolute shrinkage and selection operator (LASSO) regression, combined with fivefold cross-validation to optimize feature selection. LASSO identifies the most predictive subset by shrinking less relevant feature coefficients to zero, with the optimal set determined by the regularization parameter λ. It is important to note that all feature selection procedures (e.g., t-test, LASSO) were applied strictly within the training set, with the test set remaining completely unseen during both feature selection and model development. This approach helps prevent data leakage and ensures an unbiased evaluation. Figure 5 illustrates the experimental outcomes, including plots of feature coefficients versus the regularization parameter λ, the retained key feature coefficients, and the distribution of feature weights, thereby facilitating interpretation and analysis.Fig. 5Visualization of feature selection results using the Lasso model on the training set. (a) Mean squared error (MSE) and its variation across different λ values. (b) Changes in feature coefficients with varying λ. (c) Magnitude of feature coefficients in the Lasso model. (d) Distribution of feature weights across the model. Visualization of feature selection results using the Lasso model on the training set. (a) Mean squared error (MSE) and its variation across different λ values. (b) Changes in feature coefficients with varying λ. (c) Magnitude of feature coefficients in the Lasso model. (d) Distribution of feature weights across the model. Radiomics feature classification using machine learning models Following feature extraction and selection, the retained features are used to construct various machine learning (ML) models to perform cancer lesion differentiation, thereby evaluating the diagnostic capability of the proposed APASA strategy. In this study, five commonly used ML classifiers, spanning a variety of algorithm types, are selected to more comprehensively assess the sensitivity of different modeling ideas to the feature information acquired from the images. These ML classifiers include: logistic regression (LR), perceptron, support vector machine (SVM), K-nearest neighbor (KNN), and Naïve Bayes14,15. Among them, LR and Perceptron are linear models. LR offers greater interpretability, aligning well with clinicians’ decision-making needs, while the Perceptron model is simple, fast to train, and serves effectively as a baseline model. SVM handles high-dimensional, nonlinear data through kernel functions, providing strong generalization capability. The distance-based KNN classifier follows a lazy learning paradigm, making predictions based on instance similarity. Finally, Naïve Bayes applies Bayesian inference to update probabilities using prior knowledge, reflecting the way physicians integrate prior information with new evidence to dynamically refine their assessment of disease likelihood. All models employed grid search with 10-fold cross-validation for optimal parameter selection. This systematic evaluation covered classifiers with different levels of complexity. It allowed us to assess the robustness and generalizability of multi-regional features for benign and malignant lesion differentiation. The results also helped reveal the intrinsic discriminative value of these features. Statistical evaluation details In this study, the performance of the implemented ML models was assessed using five widely adopted classification metrics: area under the curve (AUC), accuracy (Acc), Specificity, Sensitivity, and F1-score. AUC, or the area under the receiver operating characteristic (ROC) curve, reflects a model’s overall ability to distinguish between positive and negative samples. The ROC curve is a plot that shows the performance of a classification model at various threshold settings by graphing the false positive rate (FPR) against the true positive rate (TPR). These metrics provide a comprehensive assessment of model performance. AUC reflects the overall discrimination ability, while accuracy measures the overall classification correctness. Specificity and sensitivity evaluate the ability to identify negative and positive cases, respectively. F1-score reflects the balance between precision and recall, which is particularly important for imbalanced datasets. The combined use of these five metrics enables a multifaceted evaluation of experimental results, covering global performance, class differentiation, and clinical risk control, thereby validating the effectiveness of the proposed AI-driven framework for automated cancer lesion analysis. Datasets and experimental settings Experimental data The effectiveness of the proposed framework is evaluated on two cancer types: breast and thyroid cancers. Accordingly, two distinct datasets are utilized. Ultrasound imaging is the modality of choice due to its widespread use as a primary screening tool for both breast and thyroid cancers. For breast cancer, the dataset is a private dataset acquired from Liaoning Cancer Hospital referred to as BUS-data36. For thyroid cancer, we use a publicly accessible dataset collected from Zhujiang Hospital called TNCD37. The BUS-data dataset was acquired from 300 patients between January 2023 and November 2023. For each patient video, an experienced radiologist selected the frame that provided the most informative depiction of the breast lesion. This resulted in a total of 300 images, comprising 114 benign and 186 malignant cases. The dataset was randomly split into training and test sets at an 8:2 ratio, ensuring a balanced distribution of lesion types. Specifically, the training set contained 238 cases (90 benign, 148 malignant), while the test set included 62 cases (24 benign, 38 malignant). The TNCD dataset initially comprised 3493 ultrasound images collected from 2421 patients. Images containing multiple tumors or tumors with extremely small morphology were excluded, resulting in 2500 images, of which 1375 were benign and 1125 were malignant. Then, these preselected images were randomly divided into training and test sets at an 8:2 ratio. The training set included 2000 cases (1075 benign and 925 malignant), while the test set contained 500 cases (300 benign and 200 malignant). A summary of the datasets used in the experimental validation of the proposed framework is provided in Table 2. Moreover, representative samples from both the breast cancer and thyroid cancer datasets are illustrated in Fig. 6.Table 2Dataset details for the experimental analyses of the proposed framework.Number of patientsNumber of imagesImage categoryBenignMalignantBUS-dataTotal300300114186Data splitting (8:2)Train23823890148Test62622438TNCDTotal2421250013751125Data splitting (8:2)Train–20001075925Test–500300200 Table 2Dataset details for the experimental analyses of the proposed framework.Number of patientsNumber of imagesImage categoryBenignMalignantBUS-dataTotal300300114186Data splitting (8:2)Train23823890148Test62622438TNCDTotal2421250013751125Data splitting (8:2)Train–20001075925Test–500300200 Dataset details for the experimental analyses of the proposed framework. Fig. 6Representative ultrasound images from breast and thyroid cancer datasets. Each pair shows the original image (left) and the corresponding image with the lesion highlighted (right). Representative ultrasound images from breast and thyroid cancer datasets. Each pair shows the original image (left) and the corresponding image with the lesion highlighted (right). Environmental settings All experiments in this study were conducted using the PyCharm 2024.2.1, professional edition (https://blog.jetbrains.com/pycharm/2024/08/pycharm-2024-2-1/) integrated development environment, with Python 3.9.20 (https://www.python.org/downloads/release/python-3920/) as the programming language. The feature extraction module employed the feature extractor function from the PyRadiomics 3.1.0 library (https://pypi.org/project/pyradiomics/) to obtain multiregional tumor features within a standardized radiomics analysis framework. Feature selection was performed using the LassoCV model from the scikit-learn 1.5.1 library (https://scikit-learn.cn/1.5/index.html). The experiments were run on a hardware platform equipped with the Windows 11 (25H2) operating system and an Intel Core i7-12800HX processor (2.4 GHz base frequency/4.8 GHz turbo frequency). Experimental results In this section, we evaluate the performance of our proposed APASA-based clinical decision-support system for automated cancer lesion analysis. Comparative experiments are also conducted against a conventional peritumoral feature selection method, namely morphological dilation, to highlight the advantages of our approach. As described earlier, the application of the proposed APASA leads to three ROIs: tumor ROI, peritumoral ROI, and MCG ROI. For the MCG ROI, two shape types were examined, resulting in five primary ROIs: tumor ROI, MCR ROI, MCR peritumoral ROI (MCR-peri-ROI), MCC ROI, and MCC peritumoral ROI (MCC-peri-ROI). In addition, for each MCG shape type, we also generated combined feature sets by integrating the peritumoral ROI features with those from the corresponding tumor ROI, yielding two additional cases: combined MCR (tumor ROI + MCR-peri-ROI) and combined MCC (tumor ROI + MCC-peri-ROI). Therefore, for each of the two considered cancer types, performance is evaluated under seven scenarios: tumor ROI, MCR ROI, MCR-peri-ROI, MCC ROI, MCC-peri-ROI, combined MCR, and combined MCC. Effectiveness of the proposed APASA-based framework for breast lesion differentiation Table 3 presents the breast lesion differentiation performance. For clinically significant metrics such as AUC, accuracy, and F1-score, the best and second-best results are highlighted in red and blue, respectively. Scenarios achieving a good balance between sensitivity and specificity are shown in bold. Maintaining such a balance is important, as these metrics reflect the rates of false positives and false negatives, which are critical in clinical decision-making. For all five ML models, using the combined MCR and combined MCC scenarios significantly outperformed using tumor ROI alone. The highest AUC, accuracy, and F1-score were achieved with combined MCC and combined MCR, both showing an excellent balance between sensitivity and specificity, which is critical for minimizing false positives and negatives in clinical settings. Notably, MCC ROI generally yielded better results than MCR ROI, demonstrating its suitability for this cancer lesion type. The top ML models were logistic regression for MCR and perceptron for MCC. Compared with the tumor ROI baseline, the best MCR-based model improved AUC and F1-score by 2.20% and 4.00%, respectively. The best MCC-based model achieved larger gains, with improvements of 9.64% in AUC and 17.00% in F1-score. Overall, integrating peritumoral features selected by APASA with tumor features led to substantial performance gains in breast lesion differentiation, with AUC and F1-score reaching 92.43% and 90%, respectively.Table 3Breast lesion differentiation results. Best and second-best values for AUC, accuracy, and F1-score are shown in italics and bolditalics, respectively. The best performing ML model is highlighted in boldunderline. Bolditalics indicates a good balance between sensitivity and specificity, reflecting clinical reliability.ROI selection mannerAUCAccSensitivitySpecificityF1-scoreClassifierTumor ROI89.4782.2689.4770.8382.00LR88.8282.2686.8475.0082.00SVM84.7674.1973.6875.0074.00KNN87.6183.8792.1170.8384.00Naïve Bayes82.7974.1986.8454.1773.00PerceptronMCR peritumor ROI84.4375.8178.9570.8376.00LR84.8780.6578.9583.3381.00SVM79.7874.1989.4750.0073.00KNN84.8775.8173.6879.1776.00Naïve Bayes80.4872.5876.3266.6773.00PerceptronMCR ROI selectionMCR ROI82.1361.29100.000.0047.00LR82.3574.1992.1145.8372.00SVM82.3575.8186.8458.8475.00KNN82.3572.5860.5391.6773.00Naïve Bayes67.0058.0660.5354.1759.00PerceptronCombined MCR ROI91.6785.4884.2187.5086.00LR90.4683.8781.5887.5084.00SVM88.4980.6584.2175.0081.00KNN89.8083.8786.8479.1784.00Naïve Bayes90.9085.4884.2187.5086.00PerceptronMCC peritumor ROI84.8782.2689.4770.8382.00LR84.9780.6586.8470.8380.00SVM81.3080.6589.4766.6780.00KNN85.4283.8786.8479.1784.00Naïve Bayes79.0672.5868.4279.1773.00PerceptronMCC ROISelectionMCC ROI78.0761.29100.000.0047.00LR74.2374.1989.4750.0073.00SVM81.3677.4286.8462.5077.00KNN75.7779.0386.8466.6779.00Naïve Bayes81.2561.2947.3783.3361.00PerceptronCombined MCC ROI92.3388.7192.1183.3389.00LR91.0182.2692.1166.6782.00SVM89.4785.4889.4779.1785.00KNN88.8287.1089.4783.3487.00Naïve Bayes92.4390.3294.7383.3390.00Perceptron Table 3Breast lesion differentiation results. Best and second-best values for AUC, accuracy, and F1-score are shown in italics and bolditalics, respectively. The best performing ML model is highlighted in boldunderline. Bolditalics indicates a good balance between sensitivity and specificity, reflecting clinical reliability.ROI selection mannerAUCAccSensitivitySpecificityF1-scoreClassifierTumor ROI89.4782.2689.4770.8382.00LR88.8282.2686.8475.0082.00SVM84.7674.1973.6875.0074.00KNN87.6183.8792.1170.8384.00Naïve Bayes82.7974.1986.8454.1773.00PerceptronMCR peritumor ROI84.4375.8178.9570.8376.00LR84.8780.6578.9583.3381.00SVM79.7874.1989.4750.0073.00KNN84.8775.8173.6879.1776.00Naïve Bayes80.4872.5876.3266.6773.00PerceptronMCR ROI selectionMCR ROI82.1361.29100.000.0047.00LR82.3574.1992.1145.8372.00SVM82.3575.8186.8458.8475.00KNN82.3572.5860.5391.6773.00Naïve Bayes67.0058.0660.5354.1759.00PerceptronCombined MCR ROI91.6785.4884.2187.5086.00LR90.4683.8781.5887.5084.00SVM88.4980.6584.2175.0081.00KNN89.8083.8786.8479.1784.00Naïve Bayes90.9085.4884.2187.5086.00PerceptronMCC peritumor ROI84.8782.2689.4770.8382.00LR84.9780.6586.8470.8380.00SVM81.3080.6589.4766.6780.00KNN85.4283.8786.8479.1784.00Naïve Bayes79.0672.5868.4279.1773.00PerceptronMCC ROISelectionMCC ROI78.0761.29100.000.0047.00LR74.2374.1989.4750.0073.00SVM81.3677.4286.8462.5077.00KNN75.7779.0386.8466.6779.00Naïve Bayes81.2561.2947.3783.3361.00PerceptronCombined MCC ROI92.3388.7192.1183.3389.00LR91.0182.2692.1166.6782.00SVM89.4785.4889.4779.1785.00KNN88.8287.1089.4783.3487.00Naïve Bayes92.4390.3294.7383.3390.00Perceptron Breast lesion differentiation results. Best and second-best values for AUC, accuracy, and F1-score are shown in italics and bolditalics, respectively. The best performing ML model is highlighted in boldunderline. Bolditalics indicates a good balance between sensitivity and specificity, reflecting clinical reliability. Effectiveness of the proposed APASA-based framework for thyroid lesion differentiation The lesion differentiation performance on the thyroid dataset is displayed in Table 4. Similarly, for clinically significant metrics such as AUC, accuracy, and F1 score, the best and second-best results are highlighted in red and blue, respectively. Scenarios that achieve a good balance between sensitivity and specificity are also displayed in bold. Unlike the breast lesion differentiation experiments, a different trend emerged. Although the combined MCR ROI and combined MCC ROI significantly outperformed the tumor ROI alone for all five ML models, they did not yield the best results. Except for the Perceptron classifier, whose best performance was achieved in the MCR ROI selection experiments, the highest differentiation performance was consistently obtained when using features extracted exclusively from the peritumoral region. Similar findings have been reported in previous studies38,39. Zhan et al.38 observed that, in distinguishing thyroid follicular adenomas from follicular carcinomas, the peritumor ROI-based model achieved a significantly higher AUC (81.20%) than the tumor ROI-based model (71.00%). They attributed this to the fact that pathological changes such as capsular and vascular invasion, which are more prevalent in follicular carcinomas, typically occur at the tumor periphery. Likewise, Zhang et al.39 demonstrated that malignant thyroid nodules more frequently present with a peritumoral “hard ring.” This ring was interpreted as reflecting either stromal reaction or tumor cell infiltration into surrounding tissues, resulting in tissue hardening. By contrast, hardness features within the tumor itself may be less reliable, as they can overlap between benign and malignant nodules and be further confounded by necrosis, liquefaction, or calcification. These factors may explain the relatively poor differentiation performance of ultrasound image features derived solely from the intratumor ROI. Moreover, in the MCC ROI selection experiments, although performance improved compared with using only the tumor ROI, the models still underperformed relative to the MCR ROI experiments. This may be related to the pathological characteristics of thyroid lesions. In the MCC ROI selection strategy, where the peritumoral region is defined by a minimum covering circle, a considerable amount of background tissue may be introduced, which can dilute the discriminative information and affect diagnostic performance. Furthermore, SVM was the top-performing model in both the MCR and MCC experiments. In the MCR setting, it achieved an AUC of 96.86% and an F1-score of 91.00%. In the MCC setting, it achieved an AUC of 73.28% and an F1-score of 69.00%. Overall, for both ROI shapes generated with our proposed APASA, the peritumor ROI of the MCR shape yielded the most suitable features, achieving an AUC of 96.86% and an F1-score of 91.00%. Importantly, these results reflect gains of 30.75% in AUC and 29.00% in F1-score compared with the tumor ROI baseline.Table 4Thyroid lesion differentiation results. Best and second-best values for AUC, accuracy, and F1-score are shown in italics and bolditalics, respectively. The best performing ML model is highlighted in boldunderline. Bold indicates a good balance between sensitivity and specificity, reflecting clinical reliability.ROI selection mannerAUCAccSensitivitySpecificityF1-scoreClassifierTumor ROI66.1162.2055.5066.6762.00LR66.5161.2054.5065.6761.00SVM65.5962.0053.0068.0062.00KNN69.0363.8062.5064.6764.00Naïve Bayes70.2966.0065.0066.6766.00PerceptronMCR peritumor ROI96.1990.2087.5092.0090.00LR96.8691.0096.0087.6791.00SVM94.1988.0091.0086.0088.00KNN88.1979.8089.5073.3380.00Naïve Bayes66.4562.2049.5070.6762.00PerceptronMCR ROI selectionMCR ROI73.3969.2073.5066.3370.00LR71.4968.2068.0068.3368.00SVM69.6862.6065.5060.6763.00KNN68.9662.6076.5053.3363.00Naïve Bayes62.5859.6065.5055.6760.00PerceptronCombined MCR ROI78.5070.0066.5072.3370.00LR78.6069.8067.0071.6770.00SVM85.5178.4077.0079.3379.00KNN80.6570.8079.0065.3371.00Naïve Bayes80.0971.0069.0072.3371.00PerceptronMCC peritumor ROI66.7863.8060.0066.3364.00LR65.0960.8058.0062.6761.00SVM62.6956.6058.5055.3357.00KNN66.1363.8063.5064.0064.00Naïve Bayes61.2760.2058.0061.6761.00PerceptronMCC ROI selectionMCC ROI73.7568.4066.5069.6769.00LR73.2868.8066.5070.3369.00SVM66.8063.4061.5064.6764.00KNN65.8862.0065.5059.6762.00Naïve Bayes68.4564.8067.0063.3365.00PerceptronCombined MCC ROI68.0963.6059.0066.6764.00LR68.2363.4059.0066.3364.00SVM68.7663.0060.0065.0063.00KNN70.7264.2063.5064.6765.00Naïve Bayes72.1067.0057.5073.3367.00Perceptron Table 4Thyroid lesion differentiation results. Best and second-best values for AUC, accuracy, and F1-score are shown in italics and bolditalics, respectively. The best performing ML model is highlighted in boldunderline. Bold indicates a good balance between sensitivity and specificity, reflecting clinical reliability.ROI selection mannerAUCAccSensitivitySpecificityF1-scoreClassifierTumor ROI66.1162.2055.5066.6762.00LR66.5161.2054.5065.6761.00SVM65.5962.0053.0068.0062.00KNN69.0363.8062.5064.6764.00Naïve Bayes70.2966.0065.0066.6766.00PerceptronMCR peritumor ROI96.1990.2087.5092.0090.00LR96.8691.0096.0087.6791.00SVM94.1988.0091.0086.0088.00KNN88.1979.8089.5073.3380.00Naïve Bayes66.4562.2049.5070.6762.00PerceptronMCR ROI selectionMCR ROI73.3969.2073.5066.3370.00LR71.4968.2068.0068.3368.00SVM69.6862.6065.5060.6763.00KNN68.9662.6076.5053.3363.00Naïve Bayes62.5859.6065.5055.6760.00PerceptronCombined MCR ROI78.5070.0066.5072.3370.00LR78.6069.8067.0071.6770.00SVM85.5178.4077.0079.3379.00KNN80.6570.8079.0065.3371.00Naïve Bayes80.0971.0069.0072.3371.00PerceptronMCC peritumor ROI66.7863.8060.0066.3364.00LR65.0960.8058.0062.6761.00SVM62.6956.6058.5055.3357.00KNN66.1363.8063.5064.0064.00Naïve Bayes61.2760.2058.0061.6761.00PerceptronMCC ROI selectionMCC ROI73.7568.4066.5069.6769.00LR73.2868.8066.5070.3369.00SVM66.8063.4061.5064.6764.00KNN65.8862.0065.5059.6762.00Naïve Bayes68.4564.8067.0063.3365.00PerceptronCombined MCC ROI68.0963.6059.0066.6764.00LR68.2363.4059.0066.3364.00SVM68.7663.0060.0065.0063.00KNN70.7264.2063.5064.6765.00Naïve Bayes72.1067.0057.5073.3367.00Perceptron Thyroid lesion differentiation results. Best and second-best values for AUC, accuracy, and F1-score are shown in italics and bolditalics, respectively. The best performing ML model is highlighted in boldunderline. Bold indicates a good balance between sensitivity and specificity, reflecting clinical reliability. To further evaluate the effectiveness of the proposed method, we present ROC curves of the best-performing models on the two datasets in Fig. 7. The curves demonstrate consistent diagnostic performance across different classifiers. For breast lesion differentiation, LR and Perceptron exhibit similar trends, with a rapid increase in true positive rate at low false positive rates and a gradual plateau near the upper boundary, indicating reliable discrimination across decision thresholds. For thyroid lesion differentiation, the LR and SVM curves show steeper initial rises and remain closer to the upper-left region, suggesting improved class separability. The consistency of these trends across classifiers indicates that the performance gains are primarily attributable to the informative peritumoral ROI representation generated by APASA, rather than to any specific classifier.Fig. 7(a) ROC curves of Logistic Regression and Perceptron on the breast cancer dataset. (b) ROC curves of logistic regression and SVM on the thyroid cancer dataset. (a) ROC curves of Logistic Regression and Perceptron on the breast cancer dataset. (b) ROC curves of logistic regression and SVM on the thyroid cancer dataset. Comparison with conventional peritumoral region selection methods To further highlight the advantages of our proposed APASA method over conventional approaches such as morphological dilation, we conducted additional experiments in which the peritumoral region was defined by applying a 2-pixel dilation and a 4-pixel dilation, respectively. These settings were selected as representative narrow-range expansions commonly adopted in prior studies. Existing works have explored multiple peritumoral expansion ranges and have often reported that relatively narrow regions near the tumor boundary provide more informative features than excessively enlarged regions18–20,23,24,31,33. Therefore, the chosen dilation sizes aim to reflect practical and commonly used configurations while avoiding overly large expansions that may introduce irrelevant background information. Since dilation is a commonly adopted yet coarse strategy, this comparison allows us to evaluate whether APASA offers more precise and informative region selection. The obtained results are summarized in Tables 5 and 6 for breast lesion and thyroid lesion, respectively. For breast lesions, fusing representative features from both the tumor ROI and the peritumoral ROI defined through our proposed APASA method achieves remarkably improved diagnostic performance, showing consistent gains across key evaluation metrics. For thyroid lesions, whether features are derived solely from the peritumoral ROI, the MCG ROI, or the combined MCG ROI, the APASA strategy consistently surpasses the conventional morphological dilation method by substantial margins. Overall, incorporating APASA into automated ultrasound breast or thyroid tumor lesion differentiation frameworks leads to significantly enhanced diagnostic accuracy. These results underscore its practical value and strong potential as a robust tool for adaptive appropriate ROI selection, thereby facilitating the extraction of the most informative features to achieve more precise cancer lesion analysis.Table 5Breast lesion differentiation results comparing APASA and morphological dilation for peritumoral region selection. For each ROI type, the top two AUC and F1-score values are highlighted in Bolditalics, with the corresponding machine learning model in bold.Selection manner of ROIAUCAccSensitivitySpecificityF1-scoreClassifierMCR peritumor ROI84.4375.8178.9570.8376.00LR84.8780.6578.9583.3381.00SVM79.7874.1989.4750.0073.00KNN84.8775.8173.6879.1776.00Naïve Bayes80.4872.5876.3266.6773.00PerceptronMCR based ROI selectionMCR ROI82.1361.29100.000.0047.00LR82.3574.1992.1145.8372.00SVM82.3575.8186.8458.8475.00KNN82.3572.5860.5391.6773.00Naïve Bayes67.0058.0660.5354.1759.00PerceptronCombined MCR ROI91.6785.4884.2187.5086.00LR90.4683.8781.5887.5084.00SVM88.4980.6584.2175.0081.00KNN89.8083.8786.8479.1784.00Naïve Bayes90.9085.4884.2187.5086.00PerceptronMCC peritumor ROI84.8782.2689.4770.8382.00LR84.9780.6586.8470.8380.00SVM81.3080.6589.4766.6780.00KNN85.4283.8786.8479.1784.00Naïve Bayes79.0672.5868.4279.1773.00PerceptronMCC based ROI selectionMCC ROI78.0761.29100.000.0047.00LR74.2374.1989.4750.0073.00SVM81.3677.4286.8462.5077.00KNN75.7779.0386.8466.6779.00Naïve Bayes81.2561.2947.3783.3361.00PerceptronCombined MCC ROI92.3388.7192.1183.3389.00LR91.0182.2692.1166.6782.00SVM89.4785.4889.4779.1785.00KNN88.8287.1089.4783.3487.00Naïve Bayes92.4390.3294.7383.3390.00PerceptronDilation2 peritumor ROI85.0985.4881.5891.6786.00LR84.7661.29100.000.0047.00SVM91.5682.2692.1166.6782.00KNN84.7683.8778.9591.6784.00Naïve Bayes88.9379.0376.3283.3379.00PerceptronDilation2 based ROI selectionDilation2 ROI89.6982.2686.8475.0082.00LR90.3583.8786.8479.1784.00SVM86.5180.6586.8470.8380.00KNN87.8382.2681.5883.3382.00Naïve Bayes82.2691.3481.5888.3382.00PerceptronCombined Dilation2 ROI84.7682.2681.5883.3382.00LR86.9580.6584.2175.0082.00SVM87.7782.2692.1166.6782.00KNN88.0582.2681.5883.3382.00Naïve Bayes87.2879.0381.5875.0079.00PerceptronDilation4 peritumor ROI83.9982.2676.3291.6782.00LR84.5480.6581.5879.1781.00SVM86.7979.0386.8466.6779.00KNN81.4783.8778.9491.6784.00Naïve Bayes76.4374.1971.0579.1775.00PerceptronDilation4 based ROI selectionDilation4 ROI87.0680.6586.8470.8380.00LR87.1780.6586.8470.8380.00SVM88.1679.0381.5875.0079.00KNN86.6275.8171.0583.3376.00Naïve Bayes88.0585.4884.2187.5086.00PerceptronCombined Dilation4 ROI87.2883.8778.9591.6784.00LR87.5068.3694.7429.1765.00SVM85.3679.0384.2170.8381.00KNN85.2083.8781.5887.5084.00Naïve Bayes91.2382.2684.2179.1782.00PerceptronTable 6Thyroid lesion differentiation results comparing APASA and morphological dilation for peritumoral region selection. For each ROI type, the top two AUC and F1-score values are highlighted in Bolditalics, with the corresponding machine learning model in bold.Selection manner of ROIAUCAccSensitivitySpecificityF1-scoreClassifierMCR peritumor ROI96.1990.2087.5092.0090.00LR96.8691.0096.0087.6791.00SVM94.1988.0091.0086.0088.00KNN88.1979.8089.5073.3380.00Naïve Bayes66.4562.2049.5070.6762.00PerceptronMCR ROI selectionMCR ROI73.3969.2073.5066.3370.00LR71.4968.2068.0068.3368.00SVM69.6862.6065.5060.6763.00KNN68.9662.6076.5053.3363.00Naïve Bayes62.5859.6065.5055.6760.00PerceptronCombined MCR ROI78.5070.0066.5072.3370.00LR78.6069.8067.0071.6770.00SVM85.5178.4077.0079.3379.00KNN80.6570.8079.0065.3371.00Naïve Bayes80.0971.0069.0072.3371.00PerceptronMCC peritumor ROI66.7863.8060.0066.3364.00LR65.0960.8058.0062.6761.00SVM62.6956.6058.5055.3357.00KNN66.1363.8063.5064.0064.00Naïve Bayes61.2760.2058.0061.6761.00PerceptronMCC ROI selectionMCC ROI73.7568.4066.5069.6769.00LR73.2868.8066.5070.3369.00SVM66.8063.4061.5064.6764.00KNN65.8862.0065.5059.6762.00Naïve Bayes68.4564.8067.0063.3365.00PerceptronCombined MCC ROI68.0963.6059.0066.6764.00LR68.2363.4059.0066.3364.00SVM68.7663.0060.0065.0063.00KNN70.7264.2063.5064.6765.00Naïve Bayes72.1067.0057.5073.3367.00PerceptronDilation2 peritumor ROI62.3258.2053.5061.3359.00LR61.7659.4054.5062.6760.00SVM64.2661.4057.5064.0062.00KNN63.5557.8072.5048.0058.00Naïve Bayes60.9861.6057.5064.3362.00PerceptronDilation2 based ROI selectionDilation2 ROI65.8363.6059.5066.3364.00LR65.6562.6059.0065.0063.00SVM65.8563.2058.0066.6763.00KNN67.0762.4067.5059.0063.00Naïve Bayes51.0551.4047.5054.0052.00PerceptronCombined Dilation2 ROI67.4461.2056.5064.3361.00LR67.1562.2056.5066.0062.00SVM63.5357.8051.5062.0058.00KNN69.4162.0064.5060.3362.00Naïve Bayes62.6358.8051.5063.6759.00PerceptronDilation4 peritumor ROI63.4160.6055.0064.3361.00LR63.4259.6055.5062.3360.00SVM64.0062.2054.0064.3360.00KNN66.0564.2074.0057.6764.00Naïve Bayes54.5554.4050.0057.3355.00PerceptronDilation4 based ROI selectionDilation4 ROI65.0663.0060.0065.0063.00LR67.0960.6053.0065.6761.00SVM68.3662.6061.0063.6763.00KNN62.5261.2061.5061.0062.00Naïve Bayes54.5951.4061.5044.6752.00PerceptronCombined Dilation4 ROI66.2562.0059.5063.6762.00LR66.0060.4055.5063.6761.00SVM66.0964.4057.0069.3364.00KNN70.5962.8066.5060.3363.00Naïve Bayes63.8158.2055.5060.0059.00Perceptron Table 5Breast lesion differentiation results comparing APASA and morphological dilation for peritumoral region selection. For each ROI type, the top two AUC and F1-score values are highlighted in Bolditalics, with the corresponding machine learning model in bold.Selection manner of ROIAUCAccSensitivitySpecificityF1-scoreClassifierMCR peritumor ROI84.4375.8178.9570.8376.00LR84.8780.6578.9583.3381.00SVM79.7874.1989.4750.0073.00KNN84.8775.8173.6879.1776.00Naïve Bayes80.4872.5876.3266.6773.00PerceptronMCR based ROI selectionMCR ROI82.1361.29100.000.0047.00LR82.3574.1992.1145.8372.00SVM82.3575.8186.8458.8475.00KNN82.3572.5860.5391.6773.00Naïve Bayes67.0058.0660.5354.1759.00PerceptronCombined MCR ROI91.6785.4884.2187.5086.00LR90.4683.8781.5887.5084.00SVM88.4980.6584.2175.0081.00KNN89.8083.8786.8479.1784.00Naïve Bayes90.9085.4884.2187.5086.00PerceptronMCC peritumor ROI84.8782.2689.4770.8382.00LR84.9780.6586.8470.8380.00SVM81.3080.6589.4766.6780.00KNN85.4283.8786.8479.1784.00Naïve Bayes79.0672.5868.4279.1773.00PerceptronMCC based ROI selectionMCC ROI78.0761.29100.000.0047.00LR74.2374.1989.4750.0073.00SVM81.3677.4286.8462.5077.00KNN75.7779.0386.8466.6779.00Naïve Bayes81.2561.2947.3783.3361.00PerceptronCombined MCC ROI92.3388.7192.1183.3389.00LR91.0182.2692.1166.6782.00SVM89.4785.4889.4779.1785.00KNN88.8287.1089.4783.3487.00Naïve Bayes92.4390.3294.7383.3390.00PerceptronDilation2 peritumor ROI85.0985.4881.5891.6786.00LR84.7661.29100.000.0047.00SVM91.5682.2692.1166.6782.00KNN84.7683.8778.9591.6784.00Naïve Bayes88.9379.0376.3283.3379.00PerceptronDilation2 based ROI selectionDilation2 ROI89.6982.2686.8475.0082.00LR90.3583.8786.8479.1784.00SVM86.5180.6586.8470.8380.00KNN87.8382.2681.5883.3382.00Naïve Bayes82.2691.3481.5888.3382.00PerceptronCombined Dilation2 ROI84.7682.2681.5883.3382.00LR86.9580.6584.2175.0082.00SVM87.7782.2692.1166.6782.00KNN88.0582.2681.5883.3382.00Naïve Bayes87.2879.0381.5875.0079.00PerceptronDilation4 peritumor ROI83.9982.2676.3291.6782.00LR84.5480.6581.5879.1781.00SVM86.7979.0386.8466.6779.00KNN81.4783.8778.9491.6784.00Naïve Bayes76.4374.1971.0579.1775.00PerceptronDilation4 based ROI selectionDilation4 ROI87.0680.6586.8470.8380.00LR87.1780.6586.8470.8380.00SVM88.1679.0381.5875.0079.00KNN86.6275.8171.0583.3376.00Naïve Bayes88.0585.4884.2187.5086.00PerceptronCombined Dilation4 ROI87.2883.8778.9591.6784.00LR87.5068.3694.7429.1765.00SVM85.3679.0384.2170.8381.00KNN85.2083.8781.5887.5084.00Naïve Bayes91.2382.2684.2179.1782.00Perceptron Breast lesion differentiation results comparing APASA and morphological dilation for peritumoral region selection. For each ROI type, the top two AUC and F1-score values are highlighted in Bolditalics, with the corresponding machine learning model in bold. Table 6Thyroid lesion differentiation results comparing APASA and morphological dilation for peritumoral region selection. For each ROI type, the top two AUC and F1-score values are highlighted in Bolditalics, with the corresponding machine learning model in bold.Selection manner of ROIAUCAccSensitivitySpecificityF1-scoreClassifierMCR peritumor ROI96.1990.2087.5092.0090.00LR96.8691.0096.0087.6791.00SVM94.1988.0091.0086.0088.00KNN88.1979.8089.5073.3380.00Naïve Bayes66.4562.2049.5070.6762.00PerceptronMCR ROI selectionMCR ROI73.3969.2073.5066.3370.00LR71.4968.2068.0068.3368.00SVM69.6862.6065.5060.6763.00KNN68.9662.6076.5053.3363.00Naïve Bayes62.5859.6065.5055.6760.00PerceptronCombined MCR ROI78.5070.0066.5072.3370.00LR78.6069.8067.0071.6770.00SVM85.5178.4077.0079.3379.00KNN80.6570.8079.0065.3371.00Naïve Bayes80.0971.0069.0072.3371.00PerceptronMCC peritumor ROI66.7863.8060.0066.3364.00LR65.0960.8058.0062.6761.00SVM62.6956.6058.5055.3357.00KNN66.1363.8063.5064.0064.00Naïve Bayes61.2760.2058.0061.6761.00PerceptronMCC ROI selectionMCC ROI73.7568.4066.5069.6769.00LR73.2868.8066.5070.3369.00SVM66.8063.4061.5064.6764.00KNN65.8862.0065.5059.6762.00Naïve Bayes68.4564.8067.0063.3365.00PerceptronCombined MCC ROI68.0963.6059.0066.6764.00LR68.2363.4059.0066.3364.00SVM68.7663.0060.0065.0063.00KNN70.7264.2063.5064.6765.00Naïve Bayes72.1067.0057.5073.3367.00PerceptronDilation2 peritumor ROI62.3258.2053.5061.3359.00LR61.7659.4054.5062.6760.00SVM64.2661.4057.5064.0062.00KNN63.5557.8072.5048.0058.00Naïve Bayes60.9861.6057.5064.3362.00PerceptronDilation2 based ROI selectionDilation2 ROI65.8363.6059.5066.3364.00LR65.6562.6059.0065.0063.00SVM65.8563.2058.0066.6763.00KNN67.0762.4067.5059.0063.00Naïve Bayes51.0551.4047.5054.0052.00PerceptronCombined Dilation2 ROI67.4461.2056.5064.3361.00LR67.1562.2056.5066.0062.00SVM63.5357.8051.5062.0058.00KNN69.4162.0064.5060.3362.00Naïve Bayes62.6358.8051.5063.6759.00PerceptronDilation4 peritumor ROI63.4160.6055.0064.3361.00LR63.4259.6055.5062.3360.00SVM64.0062.2054.0064.3360.00KNN66.0564.2074.0057.6764.00Naïve Bayes54.5554.4050.0057.3355.00PerceptronDilation4 based ROI selectionDilation4 ROI65.0663.0060.0065.0063.00LR67.0960.6053.0065.6761.00SVM68.3662.6061.0063.6763.00KNN62.5261.2061.5061.0062.00Naïve Bayes54.5951.4061.5044.6752.00PerceptronCombined Dilation4 ROI66.2562.0059.5063.6762.00LR66.0060.4055.5063.6761.00SVM66.0964.4057.0069.3364.00KNN70.5962.8066.5060.3363.00Naïve Bayes63.8158.2055.5060.0059.00Perceptron Thyroid lesion differentiation results comparing APASA and morphological dilation for peritumoral region selection. For each ROI type, the top two AUC and F1-score values are highlighted in Bolditalics, with the corresponding machine learning model in bold. Discussions In recent years, artificial intelligence techniques have significantly advanced medical image analysis, particularly in oncology, with strong progress in radiomic feature extraction and predictive modeling for clinical decision support. Within this context, the reliable definition of regions of interest remains a key factor influencing downstream performance. The proposed APASA method addresses this challenge through an adaptive and morphology-preserving peritumoral ROI selection strategy for ultrasound-based lesion analysis, improving the quality of feature representation for subsequent diagnostic modeling. From a methodological perspective, APASA is conceptually related to recent coverage optimization strategies proposed for sensor or access-point deployment in full line-of-sight (LoS) environments40,41. These approaches aim to eliminate redundant sensors or access points while maintaining complete coverage in complex geometries, thereby improving efficiency without compromising coverage performance. In a similar strategy, APASA seeks to reduce redundant or non-informative regions in medical images by adaptively selecting informative peritumoral areas while avoiding unnecessary inclusion of background tissue. Although the application domains differ, both approaches emphasize efficient coverage definition, redundancy reduction, and preservation of task-relevant information. The experimental results of this study demonstrate that the proposed APASA strategy improves automatic cancer lesion analysis by more effectively leveraging tumor microenvironment information through adaptive peritumoral feature extraction. This improvement can be attributed to its ability to define the peritumoral region in a morphology-agnostic and data-adaptive manner. In contrast to conventional approaches that rely on fixed morphological dilation to define the peritumoral region, APASA eliminates the need for manual tuning of dilation thickness and reduces the influence of subjective parameter selection and potential noise introduction. As a result, APASA consistently outperforms baseline methods that consider only intra-tumoral regions as well as traditional dilation-based peritumoral modeling approaches, highlighting its effectiveness as a robust ROI selection strategy for AI-driven cancer lesion analysis. Robust peritumoral selection with morphology preservation Extensive experimental analyses demonstrate that incorporating peritumoral microenvironment features identified through our proposed APASA significantly enhances diagnostic performance. Notably, these performance improvements are consistently observed across both breast and thyroid ultrasound datasets, suggesting the applicability and robustness of the approach within ultrasound-based lesion analysis. Moreover, the method is straightforward, eliminating the need for labor-intensive parameter tuning to determine pixel expansion ranges, thereby offering a more efficient and objective alternative to conventional practices. Unlike conventional morphological dilation operations, which expand lesion boundaries uniformly and risk distorting the original lesion shape, our approach selectively incorporates diagnostically relevant peritumoral regions while preserving the original lesion geometry. It is worth noting that maintaining the integrity of lesion contours is particularly important in clinical workflows, where shape-related features are crucial for malignancy assessment and decision-making. Overall, these advantages establish our proposed strategy as a practical and robust solution for integrating multi-regional features into computer-aided diagnosis systems, with strong potential for real-world clinical adoption. Optimizing peritumoral selection and model choice for different cancer lesion types The core concept of our adaptive peritumoral area selection approach is the minimum coverage graph. In this study, we implemented two MCG shapes: rectangular (MCR) and circular (MCC). Experimental results show that for the same cancer lesion type, the optimal machine learning model may differ depending on the selected MCG shape. For instance, in breast lesion differentiation, MCR yielded the highest performance with the logistic regression model, whereas MCC performed best with the perceptron. Similarly, for thyroid lesions, the best-performing models were SVM for MCR and logistic regression for MCC. These findings highlight that the geometry of the peritumoral region plays a crucial role in shaping the extracted feature profiles, which in turn influence classification performance. Rectangular and circular MCGs capture distinct spatial patterns and textural variations within the peritumoral environment, resulting in differences in the quantity and nature of radiomic features obtained. Since shape descriptors are pivotal in malignancy evaluation, such differences can substantially impact the sensitivity and specificity of predictive models. Therefore, optimizing the combination of MCG shape and machine learning model according to the specific cancer lesion type is vital for achieving superior diagnostic precision. Varying impact of peritumoral features on lesion differentiation across cancer types It is well recognized that features from the peritumoral microenvironment provide valuable complementary information for malignancy assessment and diagnostic decision-making. Traditionally, peritumoral features are viewed as n rather than primary sources of information. However, our findings and others’39 suggest that, depending on the cancer type, relying solely on peritumoral features can sometimes yield more accurate predictions than using intratumoral features alone or even their combination. This phenomenon can occur when the tumor is small or its internal characteristics do not clearly indicate malignancy or benignity. In these cases, the surrounding tissue may provide richer information that better distinguishes lesion types. Including intratumoral features alongside peritumoral ones might introduce noise or conflicting data, which can confuse the classifier and lower prediction accuracy. Furthermore, the effectiveness of combining these features depends on the specific biases and assumptions of the ML model used. Therefore, feature selection should be tailored to each cancer type. In some cases, the peritumoral microenvironment may provide the most informative features for accurate lesion differentiation. Influence of ROI shape and lesion type on model performance Our results revealed that the top-performing models varied depending on the chosen ROI (tumor, peritumor, or combined), the shape strategy (MCC or MCR), and the cancer type (breast or thyroid). For instance, while SVM achieved the highest AUC and F1 score for both MCC and MCR in certain ROI configurations, other models performed better in different combinations of ROI and shape, underscoring that no single model consistently dominated across all conditions. This variability highlights two important aspects. First, it reflects the fact that different ROI configurations capture complementary but non-identical information: tumor regions provide direct lesion-specific cues, peritumoral regions capture contextual and microenvironmental patterns, and the combined ROI integrates both. The discriminative power of each ROI may therefore align differently with the learning strengths of individual models. Second, the inconsistency across MCC and MCR suggests that the geometric representation of the region can influence feature distributions in ways that alter model suitability. These results highlight that both the ROI shape and the lesion type influence which model achieves the highest performance, emphasizing the need for careful model selection tailored to specific scenarios. Limitations and future work Despite the promising performance of our proposed APASA framework as a principled mechanism for incorporating both tumoral and peritumoral information into cancer analysis, several limitations should be acknowledged. First, the evaluation was conducted primarily on breast and thyroid tumor differentiation tasks. While these results demonstrate the method’s effectiveness in different clinical contexts, broader validation across other cancer types and larger, more diverse datasets will be necessary to further strengthen its generalizability. Second, the study relied on data collected under controlled conditions, which may not fully capture the variability present in multi-center, real-world clinical workflows. External validation on independent cohorts from multiple institutions will therefore be essential to establish clinical robustness. In addition, the proposed peritumoral region selection strategy was evaluated primarily in combination with traditional machine learning-based frameworks, and comparisons with more recent deep learning-based approaches were not extensively explored. Future work will focus on integrating the proposed APASA strategy into state-of-the-art deep learning pipelines, such as CNN-based42 or transformer-based architectures43,44, and evaluating its performance against recent deep learning methods for lesion analysis. This will further validate the generalizability and effectiveness of the proposed approach in modern AI-driven medical imaging frameworks. Furthermore, this study focused exclusively on ultrasound image data and was validated on only two datasets (breast and thyroid). As a result, claims regarding broader applicability should be considered preliminary. Extending APASA to multimodal settings, such as integrating ultrasound with elastography, CT, or MRI, may further enrich the extracted discriminative features and enhance diagnostic performance. It is also worth noting that accurate delineation of regions of interest is an important prerequisite for downstream tasks such as image synthesis and modality transformation, for example in generating synthetic CT (sCT) images from MRI using segmentation-guided approaches45. In addition, accurate and adaptive ROI definition may support radiomic-based predictive modeling in precision oncology. Imaging features are used for therapy response estimation, as demonstrated in recent PET-based studies46. This capability may further extend to broader clinical applications such as treatment response monitoring and prognosis prediction. Beyond lesion differentiation, the adaptive ROI selection strategy could potentially be applied to other medical imaging tasks where contextual information plays a critical role, such as cardiovascular abnormality detection47, pending further validation. Another limitation is that APASA relies on tumor boundary coordinates derived from manually generated tumor masks. As a result, its performance may be influenced by the quality and consistency of tumor segmentation. In the present work, robustness to segmentation variability and boundary perturbations was not explicitly evaluated. Nevertheless, APASA was assessed on two ultrasound datasets with different annotation sources and delineation procedures, providing some degree of variability in segmentation conditions. Future work will focus on conducting sensitivity analyses using perturbed masks and, where available, inter-reader segmentation variability to systematically assess the stability of the proposed approach under realistic clinical conditions. In addition, integrating APASA with automatic segmentation pipelines represents an important direction toward achieving fully automated CAD systems. Evaluating the proposed method in conjunction with state-of-the-art deep learning-based segmentation models will help assess its performance under practical deployment scenarios. Furthermore, extending validation to multimodal datasets will be essential to further examine the robustness and generalizability of the approach across different imaging conditions, thereby improving its clinical applicability. Conclusion This study introduced an adaptive peritumoral area selection approach (APASA) to address the challenge of defining informative regions of interest (ROIs) for cancer lesion analysis. By adaptively identifying surrounding regions that best capture contextual information, APASA enables the extraction of more discriminative features that enhance diagnostic outcomes. The effectiveness of the method was demonstrated through its integration into AI-based frameworks for breast and thyroid cancer lesion differentiation, where it consistently outperformed the commonly used morphological dilation operation, yielding significant performance gains across varying ROI settings. Beyond quantitative improvements, the experimental analyses provided important insights into model selection and ROI configuration, offering practical guidance for designing more clinically relevant decision support systems. By reducing reliance on heuristic ROI choices, APASA establishes a principled mechanism for incorporating both tumoral and peritumoral information into cancer diagnosis. These findings highlight the promise of our proposed strategy as a reliable tool for ROI definition, with the potential to extend to other imaging modalities and clinical applications where contextual information plays a critical role, ultimately contributing to improved clinical decision-making.