Abstract
Colorectal cancer (CRC) is the third most prevalent cancer and the second leading cause of cancer-related deaths, constituting about 10% of all cancer cases globally. T1 CRC, characterized by tumor invasion into the submucosa, represents the final stage at which curative endoscopic resection remains possible. Accurate optical diagnosis of stage T1 CRC during endoscopy is crucial as it influences treatment strategies and could lead to organ-sparing local resection instead of surgical resection. Artificial Intelligence (AI) to support endoscopic diagnosis of CRC is an upcoming field of research. However, the computer-aided diagnosis (CADx) systems that have been developed so far do not perform sufficiently and lack transparent outcomes. This study aims to enhance the explainability of AI in early CRC detection. A T1 CRC characterization system was developed and tested on endoscopic images collected from four centres, using expert-confirmed lesion characteristics. By jointly predicting clinically-grounded interpretable intermediate labels (Hiroshima classification, shape, spontaneous bleeding depression and location) alongside presence of CRC, our multi-task model based on a two-layer perceptron with attention pooling over domain-specific DINOv3 features achieves an AUC score for the T1 CRC detection of 0.81 with specificity of 0.74, precision of 0.80, and sensitivity of 0.63. This indicates effective identification of non-CRC cases while providing interpretable probability scores on endoscopic characteristics, ultimately aiming to enhance the quality of optical diagnosis and T1 CRC detection.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CaPTion_031.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Roman_Pavelkin1
BibTex
@InProceedings{CroAni_DomainSpecific_MICCAISAT2026,
author = { Croughs, Anieke AND Pavelkin, Roman AND Adams, Miranda AND Schreuder, Ramon-Michel AND Thijssen, Ayla AND Schoon, Erik J. AND van der Sommen, Fons},
title = { { Domain-Specific Clinically-Grounded Multi-Task Learning for Interpretable Early Colorectal Cancer Recognition } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17278},
month = {pending},
page = {pending}
}
