Deep learning-based segmentation of gallbladder cancer on abdominal computed tomography scans: a multicenter study

Objectives To train and validate segmentation models for automated segmentation of gallbladder cancer (GBC) lesions from contrast-enhanced CT images. Materials and methods This retrospective study comprised consecutive patients with pathologically proven treatment na & iuml;ve GBC who underwent a contrast-enhanced CT scan at four different tertiary care referral hospitals. The training and validation cohort comprised CT scans of 317 patients (center 1). The internal test cohort comprised a temporally independent cohort (n = 29) from center 1 (internal test 1). The external test cohort comprised CT scans from three centers [ (n = 85)]. We trained the state-of-the-art 2D and 3D image segmentation models, SAM Adapter, MedSAM, 3D TransUNet, SAM-Med3D, and 3D-nnU-Net, for automated segmentation of the GBC. The models' performance for GBC segmentation on the test datasets was assessed via dice score and intersection over union (IoU) using manual segmentation as the reference standard. Results The 2D models performed better than 3D models. Overall, MedSAM achieved the highest dice and IoU scores on both the internal [mean dice (SD) 0.776 (0.106) and mean IoU 0.653 (0.133)] and external [mean dice (SD) 0.763 (0.098) and mean IoU 0.637 (0.116)] test sets. Among the 3D models, TransUNet showed the best segmentation performance with mean dice (SD) and IoU (SD) of 0.479 (0.268) and 0.356 (0.235) in the internal test and 0.409 (0.339) and 0.317 (0.283) in the external test sets. The segmentation performance was not associated with GBC morphology. There was weak correlation between the dice/IoU and the size of the GBC lesions for any segmentation model. Conclusion We trained 2D and 3D GBC segmentation models on a large dataset and validated these models on external datasets. MedSAM, a 2D prompt-based foundational model, achieved the best segmentation performance.