CRC-SAM: SAM-Based Multi-Modal Segmentation and Quantification of Colorectal Cancer in CT, Colonoscopy, and Histology Images
A single model handles CT, colonoscopy, and histology images with lightweight LoRA adaptation.
Researchers have introduced CRC-SAM, a unified AI framework that segments colorectal cancer across CT scans, colonoscopy videos, and histopathology slides. Built on the MedSAM foundation model, it uses low-rank adaptation (LoRA) layers attached to a frozen encoder, allowing efficient transfer to underrepresented medical imaging modalities with minimal trainable parameters. This design keeps the model lightweight while adapting to diverse data sources.
Tested on three public datasets—MSD-Colon (CT), CVC-ClinicDB (colonoscopy), and EBHI-Seg (histology)—CRC-SAM outperformed state-of-the-art single-modality baselines across all tasks. The work, accepted as an oral presentation at ISBI 2026, demonstrates how lightweight fine-tuning of foundation models can deliver consistent, high-quality segmentation throughout the colorectal cancer diagnostic workflow, from initial imaging to pathology.
- Unified segmentation across CT, colonoscopy, and histology images using a single MedSAM-based framework
- LoRA adaptation layers enable efficient domain transfer with minimal trainable parameters
- Outperforms state-of-the-art baselines on MSD-Colon, CVC-ClinicDB, and EBHI-Seg datasets
Why It Matters
Enables consistent, automated cancer analysis across imaging modalities, potentially streamlining diagnosis and treatment planning.