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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Gallbladder cancer (GBC) diagnosis and treatment planning rely on accurate imaging.
• Automated segmentation of GBC lesions from CT scans can improve efficiency and consistency.

Purpose of the Study:

• To train and validate deep learning models for automated segmentation of GBC from contrast-enhanced CT images.
• To compare the performance of various 2D and 3D segmentation models for GBC detection.

Main Methods:

• Retrospective analysis of contrast-enhanced CT scans from 317 patients for training and validation.
• Testing of models including MedSAM, 3D TransUNet, and 3D-nnU-Net on internal (n=29) and external (n=85) datasets.
• Performance evaluation using Dice score and Intersection over Union (IoU) against manual segmentation.

Main Results:

• 2D segmentation models generally outperformed 3D models.
• MedSAM achieved the highest Dice scores (0.776 internal, 0.763 external) and IoU scores (0.653 internal, 0.637 external).
• Segmentation performance showed no association with GBC morphology or a strong correlation with lesion size.

Conclusions:

• Automated segmentation of GBC is feasible using deep learning models.
• MedSAM, a 2D prompt-based foundational model, demonstrated the best performance for GBC segmentation.
• The study validates the effectiveness of AI models on diverse, multi-center datasets.