Cascaded deep learning for automated lung cancer tumour burden quantification on [18F]FDG PET/CT Authors: Shaonan Zhong, Jie Lv, Zhaohong Pan, Youcai Li, Yimin Fu, Bingsheng Huang, Xinlu Wang Published: 2026 ABSTRACT BackgroundAccurate assessment of tumour burden in lung cancer is critical for diagnosis, prognosis, and treatment planning. To enhance segmentation accuracy and reduce false positives (FPs), we developed a cascaded deep learning framework that combines lesion segmentation and subsequent classification, aiming to enable reliable automated tumour burden estimation on positron emission tomography/computed tomography (PET/CT).MethodsIn this retrospective single-centre study, we collected 593 fluorine-18 fluorodeoxyglucose ([18F]FDG) PET/CT scans from lung cancer patients with two scanners. Scanner 1 data (N=496) were split for five-fold cross-validation (training/validation) and internal testing; Scanner 2 data (N=97) served for external testing. The proposed cascaded framework integrated a segmentation stage with a subsequent classification stage to suppress FP findings and generate lesion-level tumour segmentation for tumour burden quantification. We compared the cascaded model to standalone segmentation using Dice similarity coefficient (DSC), lesion-level precision/recall, and assessed metabolic tumour volume (MTV) and total lesion glycolysis (TLG) against manual annotations.ResultsOn internal and external test sets, the cascaded model achieved consistent segmentation accuracy (DSC =0.82) with improved precision compared to segmentation alone (P<0.05). Tumour burden estimation showed strong correlations with manual measurements (r=0.984 for MTV, r=0.998 for TLG; both P<0.05) and moderate agreement.ConclusionsThe proposed cascaded segmentation-classification architecture significantly reduces FPs and yields reliable tumour burden quantification on PET/CT, enhancing accuracy and clinical utility. Introduction Lung cancer remains one of the most common malignancies worldwide and is characterized by the fastest-growing incidence and mortality rates (1,2). The overall 5year survival rate is approximately 29%, but drops to about 10% in patients diagnosed at advanced disease (3). Medical imaging plays a pivotal role in this effort by providing comprehensive information about tumour size, location, morphology and the status of adjacent tissues (4,5). Fluorine-18 fluorodeoxyglucose ([18F]FDG) positron emission tomography/computed tomography (PET/CT) is widely regarded as a first-line, non-invasive modality for evaluating lung cancer in diverse clinical scenarios (6). Unlike single-modality techniques, PET/CT simultaneously offers metabolic and anatomical data, enabling multimodal assessment that improves diagnosis, staging and treatment planning (7). Among PET-derived quantitative biomarkers, metabolic tumour volume (MTV) and total lesion glycolysis (TLG) have demonstrated clinical utility in lung cancer staging and prognostication (8,9). These two parameters together constitute the key components of tumour burden, which reflects the overall extent and metabolic activity of malignant lesions throughout the body. Accurate delineation of target lesions is critical for external-beam radiotherapy and radiopharmaceutical therapy, where reliable tumour segmentation supports reproducible estimates of tumour burden and treatment response (10,11). Manually contouring tumours on PET/CT, however, is labour-intensive and susceptible to intra- and inter-observer variability, particularly in complex or ambiguous lesions arising from metastasis or physiological uptake (12,13). These subjective inconsistencies may contribute to suboptimal diagnostic accuracy and limit the reproducibility of tumour burden quantification (14). This underscores the clinical value of developing automated and standardized methods to reduce human variability and improve the efficiency and reliability of lung cancer assessment. Recent advances in deep learning have revitalized medical image analysis and paved the way for routine clinical decision-support tools (15-17). Convolutional neural networks, in particular, have demonstrated impressive performance in classification and segmentation tasks across multiple imaging modalities (18,19). Several groups have already explored deep learning-based analysis of [18F]FDG PET/CT in lung cancer (20-23), Wang et al. (24) developed a dual-modality PET/CT deep learning model for automated lung tumour segmentation, Park et al. (25) proposed a two-stage deep learning approach for primary lung cancer segmentation, and Carles et al. (26) reported open-source nnU-Net models for PET- and CT-based lung tumour segmentation with and without respiratory motion compensation. However, most published studies rely on relatively small cohorts, and strategies for reducing false positives (FPs), particularly those that frequently occur in PET imaging (27), remain insufficiently investigated. The lack of targeted FP suppression in automated segmentation may degrade clinical applicability and raise the FP volume—even when the Dice score appears acceptable (28). In response to this challenge, we aimed to develop an automated tumour burden quantification model for patients with lung cancer undergoing [18F]FDG PET/CT, with the intention to support nuclear medicine physicians in clinical assessment and treatment decision-making. Specifically, our study proposes a cascaded 3D segmentation-and-classification framework that suppresses FPs by feeding initial segmentation results into a dedicated classification network, thereby improving overall lesion detection. We curated a dataset of [18F]FDG PET/CT scans acquired from two different scanners. The model was trained exclusively on images from one scanner and then evaluated on datasets from both scanners to rigorously test generalizability. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0233/rc). Methods Patients We retrospectively collected preoperative [18F]FDG PET/CT imaging data and corresponding clinical information from patients with lung cancer treated at The First Affiliated Hospital of Guangzhou Medical University between 2019 and 2021. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital of Guangzhou Medical University (ES-2025-K152-01) and individual consent for this retrospective analysis was waived. A total of 593 cases were included, with 496 and 97 cases acquired from two different PET/CT scanners, respectively: GE Discovery ST 8 (Scanner 1) and GE Discovery MI (Scanner 2). Case eligibility was determined by two experienced nuclear medicine physicians through review of both the imaging studies and the associated clinical records/pathology reports. Inclusion criteria were: (I) availability of complete clinical information and histopathological confirmation of lung cancer; (II) presence of clearly visible lesions with well-defined characteristics on [18F]FDG PET/CT images; and (III) lesions exhibiting multifocal or punctate uptake patterns rather than confluent or diffusely distributed metabolic activity. Patients were excluded if they met any of the following conditions: (I) poor image quality that compromised diagnostic assessment; (II) indiscernible lesion contours on PET/CT; or (III) low-metabolic benign lesions. The overall case selection and dataset composition are summarized in Figure 1. Flowchart of dataset selection. [18F]FDG, fluorine-18 fluorodeoxyglucose; PET/CT, positron emission tomography/computed tomography. Ground truth annotation Two experienced nuclear medicine physicians (with 5 years of experience in oncologic PET/CT interpretation) collaboratively delineated all reconstructed [18F]FDG PET/CT images, with reference to available clinical diagnostic information. In cases of disagreement, consensus was reached through discussion. All annotations were performed prior to model development, and the annotators were not involved in model training or evaluation and were blinded to model outputs. Lesion segmentation was performed automatically using 3D Slicer software (version 4.11) to delineate all identifiable lesions throughout the body. Image pre-processing All PET and CT images underwent standardized preprocessing prior to model development. PET images were converted to standardized uptake value (SUV) units, and all modalities (PET, CT, and lesion masks) were resampled and co-registered to a common spatial resolution. Intensity normalization and clipping were applied to CT images to reduce non-informative background. The preprocessing pipeline was applied uniformly across datasets from both scanners. Detailed preprocessing procedures are provided in Appendix 1. Deep learning method The proposed cascaded framework comprises two stages (Figure 2): a 3D segmentation network followed by a 3D classification network. The dataset from Scanner 1 (n=496) was randomly and independently divided into training and internal testing sets using a 4:1 ratio. Of these, 400 cases were used for five-fold cross-validation during model training, while the remaining 96 cases formed the internal test set. An entire dataset from Scanner 2 (n=97) was used for external testing to evaluate generalizability. All training and evaluation stages were performed using consistent data splits. The architecture of the proposed cascade framework. 3D, three-dimensional; PET/CT, positron emission tomography/computed tomography. The first-stage segmentation model was constructed using nnU-Net-v2 (29), an open-source, self-configuring deep learning framework optimized for medical image segmentation. It has demonstrated strong performance and generalizability across various nuclear medicine segmentation applications (26,30,31). The model takes whole-body PET/CT images as input and outputs voxel-wise tumour masks. The second-stage classification model was based on classic ResNet-50 (32), which was designed to classify candidate lesions identified by the segmentation model. To generate training data, we first extracted metabolically active regions using a liver-based SUV thresholding strategy. Specifically, liver masks were obtained via the open-source TotalSegmentator tool (33), and the SUVmean and SUVstd within the liver were computed. All liver masks used for threshold estimation were visually reviewed to ensure that the reference region was free of obvious abnormal uptake or direct tumour involvement. A threshold of SUVmean + 2× SUVstd (34,35) was applied to extract candidate hypermetabolic regions across the body. Each connected component within the extracted metabolic regions was examined: regions overlapping with previously identified lesions were labelled as lesion-positive, while non-overlapping regions were labelled as lesion-negative. From each region, we extracted a 3D block that included PET, CT, and the corresponding lesion mask to serve as input for the classification network. Both the segmentation and classification networks were trained using standard configurations based on nnU-Net and ResNet frameworks. Full training details were provided in Appendix 1. During inference, the outputs of the five segmentation models trained via cross-validation were ensemble-averaged to produce the initial 3D tumour mask. Each candidate lesion—defined as a connected component in the predicted tumour mask—was then processed through an ensemble of five independently trained classification models. To further suppress FPs, a voting strategy was applied: a lesion was retained only if at least n out of 5 classifiers (where n=1, 2, or 3) classified it as a true positive (TP). Components ultimately classified as FPs were discarded, yielding the final refined tumour segmentation. Detailed descriptions of image acquisition parameters, image pre-processing procedures, and network training strategies are provided in Appendix 1. Metrics for segmentation and classification evaluation To evaluate the performance of both the initial segmentation network and the final cascaded framework in whole-body lesion identification, we employed the Dice similarity coefficient (DSC) as a metric of volumetric segmentation accuracy. Lesion-wise detection performance was assessed by determining the spatial overlap between each predicted lesion and its corresponding reference annotation. A lesion was considered a TP if it sufficiently overlapped with a ground truth lesion (overlap ratio >10%) (30); otherwise, it was classified as a false negative (FN). Conversely, any predicted lesion that did not overlap with the ground truth was regarded as a FP. Based on these definitions, we computed precision, recall, and FP per scan. FP per scan was defined as the total number of FP lesions divided by the number of scans in the corresponding dataset. Precision and recall were calculated using the following Eqs. [1–2]: To further evaluate the second-stage classification network, we calculated its sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC), providing a comprehensive assessment of its discriminatory capability in identifying TP lesions. Excluded FP lesions were retrospectively reviewed and categorized into major sources, including physiologic uptake, benign lymph node uptake, inflammatory/reactive uptake, and benign bone-related uptake. Quantification method for tumour burden To quantify tumour burden, we computed the MTV and TLG based on both the ground truth annotations and the final predicted tumour regions. Specifically, the MTV was defined as the volume of all voxels within each segmented lesion whose SUV values exceeded 40% of the lesion-specific SUVmax. These metrics were calculated using the following Eqs. [3–5], where v denotes the SUV value of each tumour voxel, and N is the total number of detected lesions: Statistical analysis Statistical analyses were conducted using commercial software (MedCalc, version 11.2.1.0; MedCalc Software). The Mann-Whitney test was used to compare baseline characteristics between different groups. The paired-sample Wilcoxon signed-rank test was applied to assess performance differences between the proposed model and the conventional model. Pearson correlation analysis and Bland-Altman plots were used to evaluate the correlation and agreement between tumour burden estimates derived from manual annotations and artificial intelligence (AI)-based segmentation. The threshold for statistical significance was set to 0.05. Results Dataset characteristics Baseline characteristics of the training, internal test, and external test cohorts are summarized (Table 1). The training set included 400 scans, while the internal and external test sets comprised 96 and 97 scans, respectively. The corresponding numbers of lesions were 663, 212, and 228. No significant differences were observed between the training and test sets in terms of age or sex. However, the external test set exhibited a distinct distribution in lesion-related characteristics, including tumour maximum diameter, volume, and SUVmax. Table 1Dataset characteristicsFeatureTraining setInternal test setExternal test setNumber of scans4009697Age (years)64 [55.8–69.2]60.5 [53.7–67.0]63.0 [54.0–68.0]Male251 (62.8)61 (63.5)64 (66.0)Total number of lesions663212228Number of lesions per scan Mean ± SD1.7±3.22.2±3.32.4±2.9 Median [IQR]1 [1–1]1 [1–2]1 [1–2]Largest tumour diameter (mm)24.7 [19.0–33.0]23.3 [16.1–34.4]17.5* [11.4–27.6]Tumour volume (cm3)4.9 [2.3–10.6]4.2 [1.7–11.5]2.3* [0.7–6.2]SUVmax6.9 [4.3–11.5]7.0 [4.6–10.8]8.1* [4.9–12.0]Values are presented as median [IQR], n (%) or mean ± SD. *, significant difference in P value when comparing with training set. IQR, interquartile range; SD, standard deviation; SUVmax, maximum standardized uptake value. Dataset characteristics Values are presented as median [IQR], n (%) or mean ± SD. *, significant difference in P value when comparing with training set. IQR, interquartile range; SD, standard deviation; SUVmax, maximum standardized uptake value. Model performance Lesion-level classification performance of the second-stage ensemble under different voting thresholds is summarized (Table 2). On the internal test set, lowering the voting threshold increased sensitivity from 0.84 (vote ≥3) to 0.91 (vote ≥2) and 0.97 (vote ≥1), while specificity decreased from 0.96 to 0.94 and 0.86; AUC remained high across thresholds (0.90–0.92). A similar trade-off was observed on the external test set, where sensitivity increased from 0.92 (vote ≥3) to 0.96 (vote ≥2) and 0.98 (vote ≥1), whereas specificity decreased from 0.94 to 0.90 and 0.73, with AUC ranging from 0.86 to 0.93. In five-fold cross-validation, the stage-2 classification network demonstrated robust discrimination, achieving a mean AUC of 0.98 with mean sensitivity and specificity of 0.92 and 0.93, respectively (Table S1). Table 2Classification performance in the proposed cascade approach (with different voting strategies)MetricVote ≥1Vote ≥2Vote ≥3Internal test AUC0.920.920.90 Sensitivity0.970.910.84 Specificity0.860.940.96External test AUC0.860.930.93 Sensitivity0.980.960.92 Specificity0.730.890.94AUC, area under the receiver operating characteristic curve. Classification performance in the proposed cascade approach (with different voting strategies) AUC, area under the receiver operating characteristic curve. Segmentation and lesion detection performance of the cascaded framework versus the baseline nnU-Net v2 is shown under different voting strategies (Table 3). As the vote ≥1 strategy achieved the highest lesion-detection sensitivity (recall), it was selected as the primary cascaded model for the subsequent quantitative analyses and visual demonstration. Under vote ≥1, the cascaded model yielded consistent performance across datasets (internal/external: DSC 0.82/0.82, precision 0.88/0.84, recall 0.84/0.84). Compared with nnU-Net v2 alone, precision rose from 0.81 to 0.88 internally and from 0.76 to 0.84 externally (both P<0.05), without a statistically significant reduction in TP detection. FP findings were significantly reduced, with FP per scan decreasing from 0.43 to 0.26 in the internal test set and from 0.62 to 0.39 in the external test set (both P<0.05). In terms of total FP lesions, the number decreased from 41 to 25 (39.0% reduction) internally and from 60 to 38 (36.7% reduction) externally. Overall segmentation performance was significantly enhanced (internal DSC improved from 0.79 to 0.82; external DSC improved from 0.80 to 0.82; both P<0.05). A similar trend was observed across five-fold cross-validation (Table S2). Table 3Segmentation and detection performance of the proposed cascade method versus baseline 3D nnUNet-v2ParametersStage 1 (3D nnUNet-v2)Proposed cascade methodVote ≥1Vote ≥2Vote ≥3Internal test DSC0.790.82*0.81*0.80* FP per scan0.430.26*0.23*0.19* Precision0.81 (178/219)0.88* (177/202)0.89* (175/197)0.91* (171/189) Recall0.84 (178/212)0.84 (177/212)0.83 (175/212)0.81 (171/212)External test DSC0.800.82*0.82*0.83* FP per scan0.620.39*0.35*0.28* Precision0.76 (193/253)0.84* (192/230)0.85* (188/222)0.87* (181/208) Recall0.85 (193/228)0.84 (192/228)0.83 (188/228)0.79* (181/228)Data in parentheses are the numerator and denominator used for calculations. *, P value indicated significant differences between 3D nnUNet-v2 and proposed method. 3D, three-dimensional; DSC, Dice similarity coefficient; FP, false positive. Segmentation and detection performance of the proposed cascade method versus baseline 3D nnUNet-v2 Data in parentheses are the numerator and denominator used for calculations. *, P value indicated significant differences between 3D nnUNet-v2 and proposed method. 3D, three-dimensional; DSC, Dice similarity coefficient; FP, false positive. Representative visual comparisons between ground truth, baseline nnU-Net v2, and the proposed cascaded model are shown for both internal and external test sets (Figure 3). Figure 4 shows that physiologic uptake accounted for the largest proportion of FP lesions excluded by the second-stage classifier (47.4%, n=18), followed by benign lymph node uptake (28.9%, n=11), benign bone-related uptake (13.2%, n=5), and inflammatory/reactive uptake (10.5%, n=4). Representative examples of these excluded FPs are provided in the Figure S1. Examples of prediction results in the internal (A) and external test (B) sets. MIP, maximum intensity projection; PET/CT, positron emission tomography/computed tomography. Composition of false-positive lesions excluded by the second-stage classifier. Tumour burden quantification Patient-level regression and concordance analyses of tumour burden are presented using the combined internal and external test sets (Figure 5). Predicted tumour burden correlated strongly with the reference standard (r=0.984 for MTV and r=0.998 for TLG; both P<0.05). Bland-Altman analysis showed small mean bias but relatively wide limits of agreement (LoA) between predicted and reference measurements. For MTV, the mean difference was 0.70 cm3 (95% LoA: −9.94 to 11.35 cm3; 96.2% within LoA) and for TLG, the mean difference was 3.04 SUV·cm3 (95% LoA: −59.07 to 65.14 SUV·cm3; 97.4% within LoA). Most cases fell within the LoA, whereas outliers were primarily associated with missed segmentation in atypical subtypes (e.g., cavitary tumours) or uncommon extrapulmonary metastases (Figure 6). Regression and concordance analysis of ground truth tumour burden versus predicted tumour burden. MTV, metabolic tumour volume; SD, standard deviation; SUV, standardized uptake value; TLG, total lesion glycolysis. Representative PET/CT fusion images illustrating typical cases of missed tumour segmentation. The red arrows indicate tumour lesions missed by the model. PET/CT, positron emission tomography/computed tomography. Discussion In this study, we developed a two-stage cascaded framework for fully automated lung cancer analysis on [18F]FDG PET/CT. By integrating a segmentation network with a dedicated classification stage, the model effectively suppressed FPs while maintaining high sensitivity. This is particularly relevant in clinical practice, where physiologic or benign FDG uptake (e.g., in inflammatory or postoperative lesions) often mimics malignancy and may lead to unnecessary follow-up investigations or invasive procedures. Compared with segmentation-only pipelines, the cascaded design significantly improved lesion-level precision and provided flexibility to balance sensitivity and specificity. Such adaptability is valuable for different clinical contexts—for example, prioritizing sensitivity in screening scenarios or emphasizing specificity in follow-up monitoring. Previous PET/CT studies have explored both segmentation-only models and multi-stage strategies for lung tumour delineation. Park et al. (25) reported a two-stage deep learning approach for [18F]FDG PET/CT lung cancer segmentation, in which the second stage was used for further segmentation refinement, whereas Carles et al. (26) developed open-source nnU-Net models for PET- and CT-based lung tumour segmentation. The reported segmentation performance of these studies was already strong, with Dice scores generally ranging from approximately 0.78 to 0.82. In contrast, our framework adopted a different second-stage design. Rather than performing an additional segmentation step, we applied lesion-level classification to candidate lesions generated from the first-stage whole-body segmentation output, with the specific aim of FP filtering rather than contour refinement. Thus, the main contribution of the present framework lies not simply in the use of a two-stage architecture, but in introducing classification-based FP suppression to support more reliable whole-body tumour burden quantification while maintaining comparable segmentation performance (DSC =0.82). Transformer-based architectures have also attracted increasing interest in medical image segmentation (36). However, directly comparable applications in whole-body [18F]FDG PET/CT tumour burden quantification remain limited, possibly in part because transformer-based models generally benefit from larger and more diverse training data or large-scale pre-training for stable optimization (37). From a practical perspective, this filtering strategy is relevant because physiologic or benign FDG uptake frequently mimics malignancy in routine PET/CT interpretation. Our additional FP analysis showed that the excluded lesions mainly consisted of physiologic and other benign or reactive findings. Moreover, the voting-based ensemble and adjustable voting threshold provide a flexible means of balancing sensitivity and precision across different clinical scenarios. The performance of AI-based segmentation models is closely tied to the size and diversity of the training data. In PET imaging, however, assembling large-scale patient cohorts is often more challenging than in other imaging modalities due to cost, access, and standardization constraints (38). In this study, we curated a dataset comprising nearly 600 patients, which represents a relatively large-scale cohort within the domain of PET-based lung cancer segmentation. Robust performance across internal cross-validation and independent test cohorts supports the generalizability of the framework to diverse scanners and acquisition protocols. This stability is essential for translation into clinical practice, where heterogeneity in imaging platforms remains a major barrier to AI deployment. Accurate and reproducible quantification of metabolic tumour burden is increasingly recognized as clinically meaningful. Metrics such as MTV and TLG from [18F]FDG-PET/CT have shown prognostic value for survival (39), utility in radiotherapy planning (40), and promise as biomarkers for immunotherapy response and patient stratification (41). In our study, automated tumour burden estimates correlated strongly with expert delineations, indicating that the framework preserved the overall trend of tumour burden well. However, Bland-Altman analysis showed only moderate agreement, with small mean bias but relatively wide LoA, suggesting that case-level discrepancies still existed in individual patients. These discrepancies were likely related to missed or imperfect segmentation in atypical lesions or uncommon metastatic sites. Therefore, the current framework may be more appropriately considered an assistive tool for tumour burden quantification rather than a fully independent replacement for manual assessment. Nevertheless, it may still offer practical clinical value by reducing manual workload and supporting more standardized evaluation in routine practice. From a practical perspective, the framework is based on widely used and reproducible deep learning components, which supports technical robustness and lowers barriers to adoption across institutions. Its modular design allows seamless integration into imaging infrastructures and decision-support systems, aligning with the increasing emphasis on workflow efficiency and standardization in nuclear medicine. This study has limitations. Rare extrapulmonary lesions were less well detected, reflecting their underrepresentation in the training set. Although data from different scanners were included, the retrospective single-centre design may still introduce bias, and clinical endpoints such as survival or treatment response were not evaluated. Future work will expand the dataset to include a broader spectrum of lung cancer subtypes and will evaluate whether automated MTV and TLG measurements can provide prognostic and predictive value in prospective trials. These efforts will be critical for demonstrating impact on patient management and supporting clinical integration. Conclusions Our proposed cascaded deep learning framework advances automated [18F]FDG PET/CT analysis by reducing FPs and delivering reliable tumour burden quantification. Its reproducibility, adaptability, and potential to enhance workflow efficiency highlight its promise as a clinically relevant tool for lung cancer assessment. Supplementary The article’s supplementary files as