Deep learning approaches for non-invasive prediction of EGFR mutation status in lung adenocarcinoma from histopathological and radiological images: a review

Abstract

Lung adenocarcinoma patients with an EGFR gene mutation respond well to targeted drugs called EGFR-TKIs. Testing for this mutation needs a tissue biopsy and lab work. This process is invasive, costly, and slow, and many patients do not have enough tissue for testing. This review asks whether deep learning can predict EGFR mutation status from images doctors already collect. These images include histopathology slides and radiology scans such as CT, PET/CT, or MRI. We searched Semantic Scholar, PubMed, Scopus, and arXiv through the Consensus platform in July 2026. The search covered studies published from 2018 to 2026. This search found 30 records, and 28 remained after removing duplicates. Nineteen studies met our criteria, reported across 20 publications. Seven studies used histopathology images, eleven used radiology images, and one combined both. Histopathology-based models reached accuracy scores (AUC) between 0.682 and 0.933. Radiology-based models reached AUC between 0.821 and 0.940. The best results came from models that combined feature types, such as radiomics with deep learning. Other strong results came from models built on large pretrained pathology foundation models, such as UNI, CONCH, and TITAN. One early study fused radiology and histopathology directly, pointing to a new multimodal direction. Most studies used heatmaps to explain their predictions, but few checked whether these explanations were reliable. Deep learning can predict EGFR mutation status from routine images. Current studies use small patient groups, and few test their models on new hospitals. Their explanations are not well validated. Larger, better-tested studies are still needed before this approach reaches clinical use.