Abstract
Colorectal cancer is the second leading cause of cancer death in the world. If the adenocarcinoma reaches the lymph nodes, it can spread and increase the risk of metastasis. Visual biomarkers that are indicative of lymph node status could be extracted and learnt using deep learning (DL). From a multi-center cohort of 393 patients, we trained and validated a DL pipeline, using 2048x2048 px patches from haematoxylin and eosin-stained whole-slide images (WSI). For each patch, a CLS token was extracted using the UNI2 foundational model, and aggregated at patient level using an attention mechanism. With focus on clinical translation, we analysed different tissue regions and attention mechanisms. The resulting models obtained better performances than health organization guidelines and indicated that (1) current histological parameters are suboptimal and (2) tumor area is not the most important tissue in predicting lymph node status. Applying foundational models to capture more specialized information allows deep learning models to learn more meaningful information and combined with an attention mechanism, showed that several parts of the WSI, and not only the tumor, are relevant for risk prediction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CaPTion_033.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=~Nil_Aren%C3%B3s_Bach1
BibTex
@InProceedings{AreNil_Attentionbased_MICCAISAT2026,
author = { Arenós Bach, Nil AND Gil, Debora AND Cano, Pau AND Musulén, Eva AND Cano, Pau},
title = { { Attention-based prediction of lymph node status in pt1 CRC using a histopathology foundational model } },
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}
}