Abstract
Pathology reports can be generated from whole-slide images (WSIs) by following a chain of thought (CoT): before writing the report, a model answers a graph of clinical questions in order. Most multiple- instance learning (MIL) models pool the image patches into a single at- tention map, then reuse that one map for every question. Methods that split attention do so across the classes of one task, never across different tasks. We remove that limit on a 93-question, 7-organ CoT challenge by giving each question its own attention pool: all 93 heads read the same frozen features, but each learns its own gated attention. Over 3 seeds, at full corpus scale and under the deployed recipe, this improves the structure of the reasoning path with no net loss of accuracy. Masking experiments show the evidence behind each question is its own: remov- ing a question’s top-attended patches lowers its predicted probability 4–5 times more, and flips its answer about 3 times more often, than removing another question’s. We use faithful in a strict sense – the model demon- strably relies on this evidence – which is weaker than matching where a pathologist looks. We also define a Specialization Index, report six stan- dard methods that did not help, and give the full development history, including a segmentation fix worth +0.123 Metric A and our official Test Phase 1 and 2 scores
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_019.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/forum?id=FQN4kYvIYs
BibTex
@InProceedings{SinGag_Faithful_MICCAISAT2026,
author = { Singh, Gagneet AND Dukre, Adinath Madhavrao AND Razzak, Imran},
title = { { Faithful Per-Task Attention for Chain-of-Thought Whole-Slide Pathology Report Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17252},
month = {pending},
page = {pending}
}