Objective To investigate the application value of a deep learning‑based computed tomography (CT) image automatic segmentation model for the clinical diagnosis of gallbladder carci-noma.

Methods The retrospective cohort study was conducted. The clinical data of 350 patients with gallbladder carcinoma who were admitted to two medical centers of Shanghai, including 114 patients in Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine and 236 patients in Renji Hospital, Shanghai Jiao Tong University School of Medicine, from January 2010 to December 2023 were collected. There were 144 males and 206 females, aged (63±11) years. All 350 patients underwent enhanced CT examination of the upper abdomen. Based on the proportion of different subtypes of gallbladder carcinoma, a stratified sampling method was employed to divide the patients into a training set of 245 cases, a validation set of 70 cases, and a test set of 35 cases in a ratio of 7∶2∶1. The training set was used to train the model, the validation set was used to optimize the model, and the test set was used to evaluate the model performance. Eight deep learning models, including convolutional neural network Unet, Unet++, ResUnet, ResUnet++ models, structural enhancement Attention‑Unet, Swin‑Unet, TransUnet‑2D model, big medical imaging MedSAM models, were selected to evaluat the automatic segmentation performance for CT image of gallbladder carcinoma. Observa-tion indicators: (1) evaluation of the automatic segmentation performance of different deep learning models for gallbladder carcinoma; (2) evaluation of the automatic segmentation performance of different deep learning models for subtypes of gallbladder carcinoma. Comparison of measurement data with normal distribution among groups was conducted using the ANOVA. Comparison of count data among groups was conducted using the chi-square test. Comparison of ordinal data among groups was conducted using the Kruskal-Wallis H test. Comparison of repeated measurement data with skewed distrubution among groups was conducted using the Friedman M test. Further pairwise comparison was conducted using Willcoxon Signed-rank test. Comparisons were adjusted using the Benjamini‑Hochberg method. The effect size of r was used to evaluate the practical significance of statistical differences.

Results (1) Evaluation of the automatic segmentation performance of different deep learning models for gallbladder carcinoma: results of the test set analysis showed that the Dice coefficients for the automatic segmentation of CT images of gallbladder carcinoma patients using the Unet, Unet++, ResUnet, ResUnet++, Attention‑Unet, Swin‑Unet, TransUnet‑2D, and MedSAM models were 0.722(0.652,0.786), 0.803(0.755,0.841), 0.756(0.694,0.808), 0.820(0.782,0.850), 0.836(0.802,0.862), 0.825(0.799,0.843), 0.828(0.798,0.850), 0.852(0.822,0.884), respectively, showing a significant difference among the 8 models (χ²=29.174, P<0.05). Further analysis showed that there were significant differences between the MedSAM model with the Unet, the Unet++, and the ResUnet model, respectively (P<0.05). The intersection-over‑union values were 0.657(0.618,0.692), 0.734(0.682,0.780), 0.707(0.640,0.768), 0.761(0.708,0.808), 0.777(0.734,0.814), 0.786(0.767,0.799), 0.775(0.755,0.789), 0.782(0.748,0.810), respectively, showing a significant difference among the 8 models (χ²=27.831, P<0.05). Further analysis showed that there were significant differences between the MedSAM model with the Unet and the ResUnet model, respectively (P<0.05). (2) Evaluation of the automatic segmentation performance of different deep learning models for subtypes of gallbladder carcinoma: results of the test set analysis showed that the Dice coefficients for the automatic segmen-tation of CT images of patients with invasive gallbladder carcinoma using the Unet, Unet++, ResUnet, ResUnet++, Attention‑Unet, Swin‑Unet, TransUnet‑2D, and MedSAM models were 0.718±0.193, 0.820±0.110, 0.739±0.160, 0.823±0.120, 0.832±0.094, 0.822±0.086, 0.826±0.083 and 0.856±0.073, respectively. For patients with mixed gallbladder carcinoma, the Dice coefficients for the automatic segmentation of CT images were 0.702±0.176, 0.779±0.134, 0.745±0.152, 0.808±0.108, 0.826±0.096, 0.820±0.073, 0.823±0.077 and 0.863±0.075, respectively. For patients with mass‑forming gallbladder carcinoma, the Dice coefficients for the automatic segmentation of CT images were 0.733±0.171, 0.802±0.125, 0.759±0.146, 0.821±0.100, 0.835±0.089, 0.828±0.070, 0.831±0.074, and 0.870±0.071, respectively. For patients with wall‑thickening gallbladder carcinoma, the Dice coefficients for the automatic segmentation of CT images were 0.710±0.195, 0.815±0.110, 0.751±0.164, 0.816±0.105, 0.831±0.092, 0.822±0.071, 0.825±0.076 and 0.865±0.076, respectively. Further visual analysis of the ResUnet++, Attention-Unet, and MedSAM models revealed ResUnet++ model showing minor missegmentation areas in invasive gallbladder carcinoma, both Attention‑Unet and ResUnet++ models showing minor missegmentation areas in wall‑thickening gallbladder carcinoma, MedSAM model showing minor missegmentation areas in mass‑forming gallbladder carcinoma and exhibiting superior segmentation results in mixed gallbladder carcinoma. Based on the Grade-CAM algorithm, MedSAM model could accurately identify and segment key structural areas of the gallbladder and tumors in various subtypes of gallbladder carcinoma, achieving high consistency between the segmentation results and the actual anatomical structures.

Conclusion Compared with traditional Unet, Unet++, and ResUnet models, the MedSAM model performs better in automatic segmentation of CT images of gallbladder carcinoma, effectively enhancing the ability to identify complex structures and lesion areas.