Abstract
In lung adenocarcinoma (LUAD), identifying actionable driver
mutations guides therapy, but next-generation sequencing(NGS) is costly,
slow, and unevenly available. Predicting mutation status directly from
routine H&E whole-slide images (WSIs) offers a rapid, low-cost com-
plement, yet prior work has focused on common drivers and largely
ignored rarer, highly actionable oncogenes. We present a data-efficient
framework that pairs frozen Prov-GigaPath foundation-model embed-
dings with lightweight classical classifiers to predict TP53, PIK3CA,
and NTRK3 status from TCGA-LUAD WSIs, and to isolate the role
of macro-scale spatial context. Using the LongNet slide encoder, which
aggregates tile embeddings with their spatial coordinates, we achieve
pooled out-of-fold AUCs of 0.666 (TP53), 0.684 (PIK3CA), and 0.688
(NTRK3), to our knowledge, the first reported image-based benchmarks
for PIK3CA and NTRK3 on TCGA-LUAD. When benchmarked against
mean-pooling and multiple-instance learning (MIL) baselines, LongNet
shows point-estimate gains on the rare oncogenes that are largest for
NTRK3, where it has a statistically significant out-performance on both
MIL baselines (DeLong p ≤ 0.011 nominal; FDR-adjusted p = 0.048).
The remaining rare-gene gains (PIK3CA and NTRK3 versus mean pool-
ing) consistently favor LongNet but did not reach significance. Further-
more, LongNet’s advantage in prediction scales inversely with mutation
frequency, as for the common driver, TP53, spatial context adds essen-
tially nothing (∆AUC +0.01 over mean pooling) but yields larger gains
for the rare oncogenes PIK3CA (+0.10) and NTRK3 (+0.11), where
position-free baselines fall toward chance. These results suggest that
foundation-model embeddings can make for a scalable molecular pre-
screening tool without needing end-to-end training.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMAI_057.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=~Kevin_Zhu3
BibTex
@InProceedings{BanAri_Predicting_MICCAISAT2026,
author = { Baniassadi, Arian AND Bezza, Thomas AND Zhu, Kevin},
title = { { Predicting Clinically Actionable LUAD Mutations from Whole-Slide Histopathology Using Prov-GigaPath Foundation Model Embeddings } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}