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

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 framework employed ResNet-50 for feature representation and integrated a gated-attention pooling mechanism for feature aggregation at the patient level. This weakly supervised approach does not require pixel-level manual annotations and enables direct slide-level prediction of drug resistance. Results The proposed gated-attention framework attained an area under the receiver operating characteristic curve (AUC) of 0.891 in 5-fold cross-validation, outperforming conventional mean-pooling, max-pooling methods, and Transformer-based architectures such as TransMIL. Ablation analyses showed that the explicit identification of discriminative histopathological areas is essential for capturing the spatial variability within tumor specimens. Using Grad-CAM visualization and HoVer-Net-based nuclear morphometric analysis, the model captured a latent morphologic risk phenotype associated with drug resistance. The resistant group exhibited higher cellular density, greater architectural disorder, greater nuclear areas (78.3 ± 22.1 µm²), higher nuclear-to-cytoplasmic ratios, and higher chromatin-texture entropy. In addition, this phenotype was closely associated with TP53 co-mutation enrichment (62.5% in the resistant group vs. 38.3% in the sensitive group), indicating underlying genomic instability. Conclusion This study demonstrates that a lightweight and weakly supervised computational-pathology framework can support the early estimation of EGFR-TKI treatment response. This pipeline is noninvasive and scalable, provides biologically explainable outputs, and may offer a practical clinical decision-support tool for precision oncology.