知镜 Cognispect · 医学文献与研究工作台

Predicting early EGFR-TKI treatment response in EGFR-mutant lung adenocarcinoma from H&E histopathology with weakly supervised multiple-instance learning

Wang W, Zhao P, Zhang Y.
Pathology, research and practice · 2026-11-01 · PMID 42981766

文献摘要

Background This study sought to construct a deep-learning pipeline using an attention mechanism and multiple-instance learning (MIL) to predict early treatment response to epidermal growth factor receptor-tyrosine kinase inhibitors (EGFR-TKIs) directly from routinely available hematoxylin-eosin (H&E) whole-slide images (WSIs) of pretreatment tissue in patients with EGFR-mutant lung adenocarcinoma (LUAD). Methods We retrospectively obtained clinicopathological data from 100 patients with EGFR-mutant lung adenocarcinoma. The model employed ResNet-50 for feature representation and integrated a gated-attention pooling mechanism for feature aggregation at the patient level. This weakly supervised approach did not require pixel-level manual annotations and enabled direct slide-level prediction of early treatment response. Results The proposed gated-attention model attained an area under the receiver operating characteristic curve (AUC) of 0.891 in 5-fold cross-validation, which was numerically higher than those of mean pooling (0.837), max pooling (0.821), ABMIL (0.869), CLAM (0.874), and TransMIL (0.863); however, the difference relative to CLAM did not reach statistical significance (unadjusted P = 0.062). In ablation analyses, replacing gated attention with mean pooling resulted in the largest reduction in AUC (0.054). Patient-level nuclear morphometric profiling revealed that the early adverse response group exhibited larger nuclear areas (78.3 ± 22.1 μm² vs. 62.7 ± 18.4 μm² in the favorable response group), along with elevated nuclear-to-cytoplasmic ratios, nuclear eccentricity, and chromatin texture entropy. Chromatin texture entropy was positively correlated with the model-predicted probability of adverse response (r = 0.58, P < 0.001). Furthermore, the adverse response group displayed a higher frequency of TP53-associated alterations (62.5% vs. 38.3%), which encompassed next-generation sequencing findings and p53 immunohistochemical surrogates. In exploratory analyses, a four-variable clinical model yielded an AUC of 0.792, whereas an integrated model incorporating the multiple instance learning (MIL) risk score demonstrated an apparent AUC of 0.918 in this cohort, warranting further out-of-sample validation. Conclusion This study demonstrates that a lightweight and weakly supervised computational-pathology model can support the early estimation of EGFR-TKI treatment response. This pipeline uses routinely acquired H&E slides without requiring additional tissue sampling, provides biologically interpretable outputs, and may support clinical decision-making in precision oncology.

在知镜精读原文来源DOI 原文

知镜提供医学文献检索、阅读笔记、人体图谱、统计绘图与研究复现工具。