Deep learning based on ultrasound for differential diagnosis of pancreatic serous cystic neoplasm and mucinous cystic neoplasm Authors: Zhiqiang Yuan, Yingyu Shi, Shuang Chen, Jiali Yang, Juan Chen, Jiawu Li, Yang Pu, Xiaojuan Ma, Qianru Yang, Xing Zhao, Lichao Mou, Yilei Shi, Yan Luo Published: 2025 ABSTRACT PurposeThis study aims to develop a deep learning (DL) model based on conventional ultrasound images for differentiating pancreatic serous neoplasms (SCN) from mucinous cystic neoplasms (MCN). We seek to determine if such a model can surpass the diagnostic performance of routine sonographic evaluation and provide a valuable, cost-effective decision-support tool.MethodsData from 459 patients with histopathologically confirmed SCN and MCN from center 1 (n = 345) and center 2 (n = 114) were retrospectively collected. Five DL models (EfficientNet-B3, Resnet-101, Densenet-121, Se-resnext-50, Inception-v3) were constructed for differential diagnosis. Simultaneously compare the diagnostic performance of DL model with that of junior radiologists (JR) and senior radiologists (SR), and determine whether DL can assist radiologists in making better diagnoses.ResultsThe EfficientNet-B3 model demonstrated the optimal predictive performance, with an area under the receiver operating curve (AUC) of 0.904 (95% CI:0.814–0.973) in the internal validation set and 0.866 (95% CI:0.783–0.936) in the external test set. In the external test set, the EfficientNet-B3 model demonstrated significantly superior diagnostic performance compared to both JR in both F1-score and specificity (all P < 0.0125), while showing fully comparable performance to both SR in these precise metrics (all P > 0.0125). With the EfficientNet-B3 model assistance, both JR showed statistically significant improvements in F1-score and specificity (all P < 0.0125), whereas both SR exhibited no statistically significant changes in either performance metric (all P > 0.0125).ConclusionsThe EfficientNet-B3 model demonstrated comparable diagnostic performance to SR while significantly enhancing the diagnostic accuracy of JR.Keywords: Ultrasound, Deep learning, Pancreatic, Serous cystic neoplasms, Mucinous cystic neoplasms, Differential diagnosis Introduction The widespread adoption of cross-sectional imaging has led to an increased detection of pancreatic cystic neoplasms (PCNs), presenting a growing diagnostic challenge. Among PCNs, serous cystic neoplasms (SCN) and mucinous cystic neoplasms (MCN) are of paramount importance due to their vastly different malignant potential and clinical management [1, 2]. While SCN are almost exclusively benign and managed with surveillance, MCN harbor a significant risk of progression to invasive carcinoma (10–39%), necessitating surgical resection [3–9]. Consequently, an accurate preoperative diagnosis is critical to avoid unnecessary surgery for SCN or dangerous delays for MCN. Deep learning (DL) has demonstrated increasing utility in pancreatic disease diagnosis, showing promising diagnostic performance [10]. Several studies have validated this approach in differentiating SCN from MCN, achieving high diagnostic accuracy using computed tomography (CT) or endoscopic ultrasound (EUS) [11–13]. For instance, a DL model applied to EUS images reported an accuracy of 82.75% [11], while another model based on CT achieved an area under the receiver operating curve (AUC) of 0.96 [13]. While existing DL models for differentiating SCN from MCN have shown promise, they are predominantly based on CT or EUS. This reliance on advanced imaging creates a translational gap with routine clinical practice, where conventional ultrasound (US) serves as the first-line imaging modality due to its accessibility and low cost. To bridge this gap, this study aims to develop and validate, to our knowledge, the first DL model based on US images, providing a practical and cost-effective tool to assist radiologists in this critical differential diagnosis. Materials and methods Patient cohort This study was approved by the Ethics Committee of West China Hospital, Sichuan University (No. 2023[2096]) and the Ethics Committee of Chengdu Shangjin Nanfu Hospital (No. 2025[0036]). Informed consent was waived for all patients due to the retrospective nature of this study. The initial patient pool was identified from the surgical pathology databases of two independent medical centers. West China Hospital of Sichuan University (center 1) contributed 1,195 patients with histopathologically confirmed SCN (n = 802) or MCN (n = 393) who underwent surgery between January 2005 and December 2024. Similarly, Shangjin Nanfu Hospital (center 2) contributed 280 patients (SCN: n = 203; MCN: n = 77) from January 2013 to December 2024. A rigorous screening process was applied to this initial cohort to ensure data quality and suitability for ultrasound-based DL analysis. The inclusion criteria were as follows: (1) confirmed by pathological examination as SCN and MCN; (2) underwent preoperative US examination within a month; and (3) no treatment was performed before examination. The exclusion criteria were as follows: (1) two tumor pathological types in the same lesion; and (2) poor quality of the ultrasound images. The baseline clinical characteristics of the patients were collected, including age, gender, clinical symptoms, tumor markers, and lesion size. Following this screening, a final cohort of 459 patients ((with one image per patient) was enrolled. The dataset comprised 345 patients from Center 1 (211 SCN and 134 MCN) and 114 patients from Center 2 (81 SCN and 33 MCN). The entire patient enrollment, exclusion, and dataset partition workflow is detailed in Fig. 1. Flowchart of patient selection and dataset partition in the study. SCN, serous cystic neoplasms; MCN, mucinous cystic neoplasms; Center 1, West China Hospital of Sichuan University; Center 2, Shangjin Nanfu Hospital Conventional ultrasound examination Patients were fasted for more than 8 h and the pancreas was adequately scanned in the supine or lateral position. The size, margin, shape and internal components of the lesions were carefully observed, and the ultrasonographic images with typical characteristics were obtained. All images analyzed in this study were conventional grayscale transabdominal ultrasound scans, acquired using equipment that included the iU22 and EPIQ7 (Philips, Amsterdam, the Netherlands), and the Resona 7 and Resona 9 (Mindray, Shenzhen, China). Data preprocess This study employed stratified random sampling (stratified by study center and pathological type), with a total of 20% cases (n = 92) selected. This sample size ensured both subgroup representation and met the minimum requirements for statistical analysis in evaluating inter-observer segmentation consistency. Two abdominal radiologists (with 4 and 6 years of abdominal ultrasound experience, respectively) independently performed manual lesion segmentation on all sampled cases, with consistency quantified using Dice Similarity Coefficient (DSC) and Intersection-over-Union (IOU) metrics. The remaining cases were segmented by the more experienced radiologist (6 years’ experience). All segmentation results underwent blinded quality review by a senior imaging expert (15 years’ abdominal ultrasound experience), with manual corrections applied to discordant regions to ensure annotation quality. To improve the accuracy of DL models in identifying pancreatic tumors, we extract key regions based on the lesion contours annotated by radiologists. The preprocessing pipeline consisted of three sequential steps. First, the minimum bounding rectangle encompassing the lesion, as delineated by radiologist annotations, was identified and cropped from the original ultrasound image. Second, the resulting rectangular image patch was converted into a square by applying zero-padding symmetrically along the shorter edges. Finally, all square patches were resized to a uniform resolution of 224 × 224 pixels to serve as standardized inputs for the subsequent convolutional neural network (CNN). This process optimizes the pancreatic ultrasound data by reducing interference from surrounding areas, thereby enhancing the model’s ability to extract features from the lesion ultrasound images. Given the limited amount of data we have collected, to reduce the risk of model overfitting and enhance its generalization ability, data augmentation techniques are applied during the training of the DL model to increase the scale and diversity of the data. These techniques include mirror flipping, rotation, adding noise, random brightness adjustment, random contrast adjustment, and color jittering. It is important to emphasize that, to objectively evaluate the model’s performance on the validation set and the independent test set, data augmentation is not applied to these two datasets, ensuring the accuracy of model evaluation. Before inputting the data into the model, all images undergo mean normalization, as described by the following formula: Here, x represents the original pixel value, µ is the mean of all pixel values, σ is the standard deviation, and xnormalized ​is the normalized pixel value. DL model construction In this study, we trained several classic CNN architectures, including Inception [14], ResNet [15], DenseNet [16], EfficientNet [17], and ResNeXt [18]. Layers are the functional units of a CNN, where each layer learns features from the input data and stores parameters. The number of layers directly affects the model’s size, computational speed, feature extraction capability, and performance. Through experimentation, we identified the optimal subtypes of each CNN architecture for our pancreatic ultrasound dataset, namely EfficientNet-B3, Resnet-101, Densenet-121, Se-resnext-50, Inception-v3. To accelerate model convergence, we employed transfer learning [19] during the training process, where the models were first pre-trained on the ImageNet dataset [20] and then fine-tuned on our pancreatic ultrasound dataset. The final fully connected layer of the model outputs two values, which are processed by the SoftMax activation function to represent the probabilities of the input pancreatic lesion being SCN or MCN, respectively. The model’s output is determined by the higher probability. The workflow of the classification model is illustrated in Fig. 2. During the training of the classification model, the cross-entropy loss function and the Adam optimizer were used to optimize the model parameters, with a batch size of 32 and 1000 epochs. The initial learning rate was set to 0.001, and it was halved if no loss reduction was observed on the validation set for 30 consecutive epochs. For all CNN architectures, the model with the lowest validation loss was selected for performance comparison. All experimental models were built using the PyTorch (version 1.10.1) DL framework based on Python (version 3.6.13) and were trained and tested on a server equipped with an NVIDIA RTX 4090 GPU to accelerate computation. To better understand the model’s predictive outcomes, a qualitative evaluation of the diagnostic approach was performed using Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm to generate heatmaps [21]. The workflow of the classification model. SCN, serous cystic neoplasm; MCN, mucinous cystic neoplasm Radiologists’ performance with and without DL model assistance Four radiologists (two junior and two senior) independently evaluated the external test set in two phases: First, they performed initial diagnoses without access to clinical information or other imaging examinations. Subsequently, after a 4-week interval to minimize recall bias, they reassessed the same cases with the assistance of the DL model. Diagnostic performance metrics were compared between unassisted and DL-assisted interpretations. Statistical analysis Statistical analysis was performed using Python (version 3.12.5) and SPSS (version 27.0). For comparison of clinical baseline characteristics, continuous variables with normal distribution were analyzed using independent samples t-tests, while non-normally distributed continuous variables were assessed with Mann-Whitney U tests. Categorical variables were compared using chi-square tests, with a two-sided significance level of α = 0.05. The differences (Δ) in diagnostic performance metrics (F1-score and specificity) for the two key comparisons—(1) between the DL model and radiologists, and (2) between unassisted and DL-assisted interpretations for each radiologist—were calculated using nonparametric bootstrap resampling (10,000 iterations) with 95% confidence intervals (CI). To account for multiple comparisons, the Bonferroni correction was applied within two independent families of hypothesis tests: the significance threshold was adjusted to P < 0.0125 (0.05 / 4) for the four comparisons between the model and the four radiologists, and similarly for the four comparisons of each radiologist’s performance with and without model assistance. A difference was considered statistically significant only if it remained significant after this family-wise correction (i.e., P < 0.0125 and the 95% bootstrap CI excluded zero). Results Clinical baseline characteristics This study retrospectively enrolled a total of 459 patients, including 292 cases of SCN and 167 cases of MCN. The baseline clinical characteristics of the patients, such as age, sex, lesion size, clinical symptoms, and tumor markers (CEA, CA19-9, and CA125), are presented in Table 1; tumor marker data were incomplete for a subset of patients. Lesion size showed a statistically significant difference between the two groups across all sets. Table 1Baseline clinical features of the datasetsClinical featuresTraining set (n = 276)Validation set (n = 69)External-test set (n = 114)SCN (n = 169)MCN (n = 107) P SCN (n = 42)MCN (n = 27) P SCN (n = 81)MCN (n = 33) P Age (year)48.7 ± 12.947.6 ± 12.50.47245.1 ± 12.846.9 ± 18.30.64148.1 ± 13.647.2 ± 13.90.755Gender0.333>0.990.215 Male331674184 Female1369135236329Size (cm)3.0(2.1,3.9)4.5(3.3,6.5)< 0.0012.8(2.0,4.2)4.9(3.0,5.8)0.0013.0(2.4,4.9)5.2(3.9,6.2)< 0.001Symptoms0.0240.0040.065 No93442995215 Yes766313182918CEA (ng/ml)0.2940.092>0.99 <51629442227432 ≥5130320CA199 (U/ml)< 0.0010.0850.156 <301537336177327 ≥309246845CA125 (U/ml)0.2060.2180.649 <241538338186126 ≥24462492Continuous variables are expressed as mean ± standard deviation or median (interquartile range)SCN serous cystic neoplasm, MCN mucinous cystic neoplasm Baseline clinical features of the datasets Continuous variables are expressed as mean ± standard deviation or median (interquartile range) SCN serous cystic neoplasm, MCN mucinous cystic neoplasm Results of the consistency evaluation of lesion contour delineation by two radiologists In the evaluation of the consistency of lesion contour delineation between the two radiologists, the two radiologists showed good consistency. The overall Intersection over Union (IOU) was 0.853 (0.783, 0.892), and the Dice Similarity Coefficient (DSC) was 0.920 (0.878, 0.943). Diagnostic Performance of the DL model This study evaluated the diagnostic performance of five DL models in differentiating SCN and MCN (Fig. 3; Table 2). The results demonstrated that the EfficientNet-B3 model exhibited superior diagnostic performance in both the internal validation set and external test set. Specifically, in the internal validation set, the model achieved an AUC of 0.904, accuracy of 0.870, F1-score of 0.836, sensitivity of 0.852, specificity of 0.881, PPV of 0.821, and NPV of 0.902. In the external test set, it showed an AUC of 0.866, accuracy of 0.807, F1-score of 0.694, sensitivity of 0.758, specificity of 0.828, PPV of 0.641, and NPV of 0.893. The representative heatmaps generated by the EfficientNet-B3 classification model for differential diagnosis between the two diseases are presented in Fig. 4. It can be seen from the superimposed heatmap that the regions on which the model bases its diagnosis are highly correlated with the lesions in the image, and are consistent with the regions that the radiologists focus on. This indicates that the model is capable of paying attention to the locations of the lesion features in the image. Receiver operating characteristic curves of different deep learning models in the validation set (a) and external-test set (b) Table 2The diagnostic performance of different DL modelsModelData SetAUCAccuracyF1 ScoreSensitivitySpecificityPPVNPVEfficientNet-B3Validation set0.904[0.814–0.973]0.8700.8360.8520.8810.8210.902External test set0.866[0.783–0.936]0.8070.6940.7580.8280.6410.893Resnet-101Validation set0.896[0.811–0.963]0.8410.8140.8890.8100.7500.919External test set0.852[0.775–0.920]0.7890.6760.7580.8000.6100.890Densenet-121Validation set0.901[0.813–0.972]0.8260.7860.8150.8330.7590.875External test set0.848[0.773–0.915]0.7890.6920.8180.7780.6000.913Se-resnext-50Validation set0.848[0.740–0.941]0.8410.8070.8520.8330.7670.897External test set0.832[0.752–0.902]0.7540.6590.8180.7280.5510.908Inception-v3Validation set0.861[0.766–0.941]0.7970.7590.8150.7860.7100.868External test set0.862[0.789–0.924]0.7900.6840.7880.7900.6050.901AUC Area Under the Curve, PPV positive predictive value, NPV negative predictive value The diagnostic performance of different DL models AUC Area Under the Curve, PPV positive predictive value, NPV negative predictive value Generate representative heatmaps using the Grad-CAM algorithm. The red areas on the heatmap indicate that the pixel points of the image in these locations play a crucial role in the model’s judgment process. a/b – mucinous cystic neoplasm, c/d – serous cystic neoplasm Diagnostic performance: DL model versus radiologists The diagnostic performance of two junior radiologists (JR1, JR2), two senior radiologists (SR1, SR2), and the EfficientNet-B3 model in the external test set is presented in Table 3. The model significantly outperformed both JR in both F1-score and specificity: versus JR1 (ΔF1 = + 0.257, 95% CI 0.141 to 0.377, P < 0.001; ΔSpecificity = + 0.211, 95% CI 0.118 to 0.306, P < 0.001) and versus JR2 (ΔF1 = + 0.264, 95% CI 0.153 to 0.378, P < 0.001; ΔSpecificity = + 0.273, 95% CI 0.171 to 0.381, P < 0.001). Conversely, its performance was on par with the senior radiologists, with no significant differences in F1-score (vs. SR1: Δ = +0.010, 95% CI -0.088 to 0.110, P = 0.428; vs. SR2: Δ = +0.038, 95% CI -0.061 to 0.143, P = 0.236) or specificity (vs. SR1: Δ = +0.013, 95% CI -0.066 to 0.094, P = 0.437; vs. SR2: Δ = +0.004, 95% CI -0.084 to 0.086, P = 0.555). Table 3The diagnostic performance of the model and radiologists in the external-testReader / ModelAccuracyF1 ScoreSpecificitySensitivityPPVNPVEfficientNet-B30.8070.6940.8280.7580.6410.893JR10.5960.4390.6170.5450.3670.769JR20.5610.4320.5560.5760.3450.763SR10.7980.6850.8150.7580.6250.892SR20.7890.6570.8270.6970.6220.870PPV Positive predictive value, NPV Negative predictive value, JR junior radiologist, SR senior radiologist The diagnostic performance of the model and radiologists in the external-test PPV Positive predictive value, NPV Negative predictive value, JR junior radiologist, SR senior radiologist Radiologists’ diagnostic performance: with versus without DL model assistance The diagnostic performance of JR and SR with and without EfficientNet-B3 model assistance in the external-test set is presented in Fig. 5; Table 4. With the assistance of the EfficientNet-B3 model, both JR showed significant improvements: JR1 achieved a ΔF1-score of + 0.156 (95% CI, 0.058 to 0.258; P < 0.001) and a Δspecificity of + 0.149 (95% CI, 0.067 to 0.235; P < 0.001), while JR2 demonstrated a ΔF1-score of + 0.166 (95% CI, 0.057 to 0.279; P = 0.002) and a Δspecificity of + 0.186 (95% CI, 0.099 to 0.280; P < 0.001).In contrast, the SR exhibited no statistically significant changes. For SR1, the ΔF1-score was − 0.009 (95% CI, − 0.032 to 0; P > 0.99) and the Δspecificity was − 0.012 (95% CI, − 0.040 to 0; P > 0.99). For SR2, the ΔF1-score was + 0.010 (95% CI, 0 to 0.032; P = 0.36) and the Δspecificity was + 0.012 (95% CI, 0 to 0.040; P = 0.36). Radiologists’ diagnostic performance without/with EfficientNet-B3 model assistance in the external-test set Table 4The diagnostic performance of radiologists without/with EfficientNet-B3 model assistance in the external-test setAccuracyF1 ScoreSpecificitySensitivityPPVNPVJR1 without assistance0.5960.4390.6170.5450.3670.769JR1 with assistance0.7370.5950.7650.6670.5370.849JR2 without assistance0.5610.4320.5560.5760.3450.763JR2 with assistance0.7280.5970.7410.6970.5230.857SR1 without assistance0.7980.6850.8150.7580.6250.892SR1 with assistance0.7890.6760.8020.7580.6100.890SR2 without assistance0.7890.6570.8270.6970.6220.870SR2 with assistance0.7980.6670.8400.6970.6390.872PPV Positive predictive value, NPV Negative predictive value, JR junior radiologist, SR senior radiologist The diagnostic performance of radiologists without/with EfficientNet-B3 model assistance in the external-test set PPV Positive predictive value, NPV Negative predictive value, JR junior radiologist, SR senior radiologist Discussion This study developed five DL models (EfficientNet-B3, Resnet-101, Densenet-121, Se-resnext-50, Inception-v3) based on US for differentiating SCN from MCN. Among these, the EfficientNet-B3 model demonstrated best diagnostic performance in both the validation set (AUC: 0.904) and external test set (AUC: 0.866). In the external test set, the diagnostic performance of the EfficientNet-B3 model, as measured by the F1 score and specificity, is significantly better than that of JR and comparable to that of SR. DL model substantially augmented diagnostic accuracy among JR - suggesting its value as a training aid. However, with the assistance of the DL model, the diagnostic performance of SR has not significantly improved. This is related to the fact that they possess a mature and precise diagnostic thinking system, and their diagnostic level has approached or reached a relatively high standard in this field. The biological behaviors of pancreatic SCN and MCN differ, necessitating distinct clinical management strategies. According to current guidelines [9], surgical resection is recommended for MCN meeting any of the following criteria: (1) size ≥ 40 mm, (2) presence of symptoms, or (3) existence of mural nodules (regardless of size). For asymptomatic MCN < 40 mm without mural nodules, surveillance with imaging every 6 months during the first year is advised, transitioning to annual follow-up if stable. Regarding SCN, surgical intervention is indicated only for symptomatic cases demonstrating mass effect. Asymptomatic SCN with radiological confirmation should undergo 1-year imaging surveillance, followed by symptom-based monitoring thereafter. Therefore, accurate differential diagnosis is of paramount clinical significance. DL-based CNN have achieved groundbreaking advances in image classification and diagnostic tasks [22–31]. DL has been extensively applied to pancreatic tasks including organ localization, volumetric segmentation, lesion detection, and diagnostic classification [32–39], with consistently satisfactory outcomes across these applications. Existing research on DL applications for pancreatic tumor diagnosis has primarily focused on detection of pancreatic adenocarcinoma [10, 35, 36, 40, 41]. Current studies investigating DL for differentiating pancreatic SCN from MCN are predominantly based on CT, EUS, and Magnetic Resonance Imaging (MRI) [11, 42–44]. US is currently the first-line imaging modality for pancreatic diseases due to its accessibility and cost-effectiveness. To date, no published studies have investigated US-based DL models for differentiating pancreatic SCN from MCN, highlighting the significant clinical value of developing such an approach. This study developed five CNN-based DL models for differentiating pancreatic SCN from MCN, among which the EfficientNet-B3 architecture demonstrated superior diagnostic performance in both the validation and external test sets. The DL model demonstrated comparable diagnostic performance to SR while significantly enhancing the diagnostic accuracy of JR. The heatmaps generated by EfficientNet-B3 confirm that the model’s areas of focus are in line with the diagnostic reasoning of radiologists, as it consistently attends to clinically critical features such as internal texture and lesion margins. This congruence demonstrates that the model has learned to emulate a human expert’s analytical approach, thereby validating its clinical applicability and fostering essential trust. Future work will focus on the quantitative validation of these attention maps to further strengthen the interpretability claims. This study has several limitations. First, as the dataset comprised only surgically resected cases, the model’s generalizability to small, incidental SCN not requiring surgery requires further validation. Second, the inherent class imbalance may introduce bias, which we mitigated with a weighted loss function and comprehensive evaluation metrics (e.g., AUC, Sensitivity), though it may still affect decision calibration [45]. Furthermore, variations in ultrasound systems and operators across centers could influence performance, despite our standardized image preprocessing and augmentation strategies. Finally, the model currently uses only grayscale images and would benefit from incorporating multimodal data. Future work will involve multi-center collaborations to build more representative datasets that include non-surgical cases, enabling a precise evaluation of the model’s clinical utility, and the development of multi-modal fusion models that integrate additional information to improve diagnostic performance. Conclusion The diagnostic performance of the EfficientNet-B3 model is comparable to that of SR and significantly superior to JR. With the assistance of the model, the diagnostic performance of JR can be significantly improved, facilitating accurate differential diagnosis between SCN and MCN. Acknowledgements We acknowledge Xing Zhao, Lichao Mou, and Yilei Shi for their valuable contributions to the model development process. Authors’ contributions Zhiqiang Yuan: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Software; Visualization; Writing-Original Draft. Yingyu Shi and Shuang Chen: Data curation; Formal analysis; Supervision; Validation. Jiali Yang: and Juan Chen: Formal analysis; Investigation. Jiawu Li and Yang Pu: Formal analysis; Data curation. Xiaojuan Ma and Qianru Yang: Resources; Supervision; Validation. Xing Zhao and Lichao Mou: Data curation; Methodology; Software; Visualization; Data curation; Methodology; Validation. Yilei Shi: Project administration; Supervision; Validation; Writing-Review & Editing. Yan Luo: Conceptualization; Data curation; Funding acquisition; Project administration; Supervision; Writing-Review & Editing. Funding This study has received funding by the National Natural Science Foundation of China (No. 82371977), the Sichuan Science and Technology Program(No. 2023NSFSC1712), and the Doctoral Research Foundation of West China Hospital, Sichuan University (No. 21hxbh039). Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate All procedural aspects of this investigation were conducted in accordance with the 2013 revision of the Declaration of Helsinki. The study was approved by the Ethics Committee of West China Hospital, Sichuan University (No. 2023[2096]) and the Ethics Committee of Chengdu Shangjin Nanfu Hospital (No. 2025[0036]). Informed consent was waived for all patients due to the retrospective nature of this study. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Yilei Shi, Email: yilei.shi@medimagingai.com. Yan Luo, Email: yanluo@scu.edu.cn. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.