Abstract
Slide-level prediction from H&E whole-slide images (WSIs)
with attention-based multiple-instance learning (MIL) is costly because
foundation model features must be extracted for every tile, a cost regardless
of how a model is deployed. We investigate how many tiles are
required for accurate predictions and design an efficient tile selector that
is able to determine the most important tiles using just a thumbnail
image. On three tasks of increasing difficulty (breast cancer ER status,
ovarian and renal subtyping) we sweep tile budgets from 100% down
to a single tile, comparing random selection with a lightweight U-Net
that predicts the attention map from the slide thumbnail. The U-Net
preserves full-bag performance (AUROC 0:79 and 0:97; Cohen’s 0:96)
down to 0:25% of tiles and, in the tail, matches or exceeds ranking by
the classifier’s own attention (which, unlike the thumbnail selector, must
encode every tile). Because this selector reads only the thumbnail, it encodes
features for the few retained tiles rather than all of them, reducing
per-slide compute, storage, and latency by about 200 (46 vs. 9;300 tiles;
0:37 s vs. 75 s) relative to full-bag inference. On the most separable task a
learned selector cuts up to 25 deeper than random before performance
drops; on the hardest task (ER status) all selectors stay close together
until the last few tiles, and the advantage there is modest. A thumbnaildriven
selector thus makes attention-MIL inference substantially cheaper
to run.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COMPAYL_065.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/COMPAYL_065_supp.pdf
Link to Open Review
BibTex
@InProceedings{PulDag_ThumbnailBased_MICCAISAT2026,
author = { Pulido Arias, Dagoberto AND Cleveland, Mason AND Mindroc-Filimon, Diana AND Kim, Albert AND Bridge, Christopher P.},
title = { { Thumbnail-Based Tile Selection for Efficient Whole-Slide Image Classification in Resource-Limited Pathology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17251},
month = {pending},
page = {pending}
}