Deep-Learning Model Drives Histopathology-Based Biomarker Detection in NSCLC

Deep-learning algorithms used as AI classifiers showed potential for detecting the presence and absence of biomarkers in tissue samples from patients with non–small cell lung cancer. These findings, published in npj Precision Oncology, support the use of deep-learning tools for more informed clinical decision-making.

To address gaps in molecular testing rates and approaches in patients with lung cancers, researchers sought newer tools to broaden patient access to biomarker testing to guide treatment decision-making in lung cancer. Analysis of histology images could allow for earlier detection of genomic alterations in tissue samples, and the application of AI tools could simplify the detection of biomarkers from slide images, as has been seen in prior studies.

Previously, the team explored the use of deep learning to aid in biomarker detection in a single-site cohort of non–small cell lung cancer slides. In this study, the researchers expanded their data set and developed deep-learning models.

The AI algorithms achieved areas under the curve of 0.87 for the identification of EGFR alterations, 0.96 for ALK, 0.88 for BRAF, and 0.83 for MET. The EGFR classifier showed greater accuracy for more common mutations, such as exon 19 deletions and L858R, than uncommon mutations, such as exon 20 insertions (area under the curve = 0.89 vs 0.81).

High accuracy was also achieved for the identification of cases that did not have any alterations. The negative predictive value was 98.6% for the EGFR classifier, 99.7% for the ALK classifier, 100% for the BRAF classifier, and 99.6% for the MET classifier. False positives were more commonly seen than false negatives, though the ranges varied widely by gene.

The study authors recommended prospective clinical study of AI classifiers before this approach could be used in real-world clinical practice.

Model Methods

To develop biomarker-classifiers, researchers used the digital pathology foundation model CanvOI 1.1, which was pretrained on a vast quantity of unlabeled histological images to identify complex patterns and features. Additionally, task-specific multiple instance learning models were trained on a cohort of tissue samples from patients with non–small cell lung cancer from various sources. The researchers cross-validated the biomarker classifiers on a validation set of 3,997 whole-slide images with a leave-one-group-out strategy for greater generalizability.

The resulting AI classifiers for EGFR, ALK, BRAF, and MET alterations were tested on hematoxylin and eosin (H&E)–stained tissue in a cohort of 968 samples from patients with non–small cell lung cancer collected from Sheba Medical Center in Israel and Hoag Memorial Hospital Presbyterian in the United States.

“Integrating AI biomarker classifiers early in the diagnostic process has the potential to provide timely and relevant insights into the likelihood of the presence or absence of targetable alterations. These tools could prioritize subsequent testing, streamline time management, optimize tissue use and reduce the need for additional procedures,” the study authors, including corresponding author Christian Rolfo, MD, PhD, Director, Division of Medical Oncology, The Arthur G. James Comprehensive Cancer Center, The Ohio State University, wrote in their report. “Moreover, a rapid and accessible tool for identifying actionable oncogenic driver mutations in NSCLC could significantly impact therapy selection, leading to better clinical outcomes by enabling treatments tailored to specific molecular profiles.”

DISCLOSURE: The study was funded by Imagene. For full author disclosures, visit nature.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.