Region-specific deep learning models for accurate segmentation of rectal structures on post-chemoradiation T2w MRI: a...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Accurate radiological evaluation of rectal cancer extent and regression post-neoadjuvant therapy is crucial.
• Manual segmentation of rectal structures on MRI is time-consuming and prone to variability.
• Computational approaches like radiomics require precise annotations of rectal wall, lumen, and perirectal fat.

Purpose of the Study:

• To develop and evaluate region-specific U-Net deep learning models for automatic segmentation of rectal structures.
• To assess the performance of these models on post-treatment T2-weighted MRI scans.

Main Methods:

• Application of U-Net deep learning models with region-specific context for segmentation.
• Training and testing on T2-weighted MRI scans from multiple institutions.
• Comparison of region-specific U-Nets against multi-reader performance and a multi-class U-Net.

Main Results:

• Region-specific U-Nets achieved high Dice scores comparable to human readers for wall and lumen segmentation.
• These models demonstrated a 20% improvement in segmentation accuracy compared to a multi-class U-Net.
• Performance remained robust even on scans with lower image quality or from external institutions.

Conclusions:

• Deep learning segmentation models with region-specific context enable accurate, detailed annotations of rectal structures on post-chemoradiation MRI.
• This facilitates improved in vivo evaluation of tumor extent and the development of image-based analytic tools for rectal cancer.