Tuning vision foundation models for rectal cancer segmentation from CT scans

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Accurate rectal cancer segmentation in CT scans is vital for diagnosis.
• Challenges exist due to complex rectal anatomy and limited annotated datasets.

Purpose of the Study:

• To develop and validate a novel segmentation model for rectal cancer.
• To create a comprehensive annotated dataset for rectal cancer research.

Main Methods:

• Introduced the U-SAM model incorporating prompt information for complex anatomy.
• Utilized the CARE Dataset with 33,024 slice pairs from 398 patients.
• Evaluated performance using Dice Coefficient, IoU, and NSD, and conducted an observer study.

Main Results:

• U-SAM achieved Dice scores of 71.23% (normal) and 76.38% (tumor), outperforming state-of-the-art methods.
• The model demonstrated diagnostic results comparable to experienced clinicians in rapid inference times.
• Observer study confirmed clinical applicability and efficiency.

Conclusions:

• U-SAM provides an efficient and reliable solution for rectal cancer and normal tissue segmentation.
• The model significantly reduces diagnostic time and assists radiologists.
• This work is expected to advance future CT-based rectal cancer diagnosis.