From pixel to precision: A systematic review of CT radiomics in the diagnosis of the <i>EGFR</i> mutation in non–small cell lung carcinoma.

Abstract

e20001 Background: Lung cancer is the leading cause of cancer mortality, with non-small cell lung cancer being more common. Radiomics is a non-invasive approach to detect actionable mutations for targeted therapy, such as EGFR in NSCLC, as current techniques like tissue biopsy and liquid biopsy are invasive and not always feasible. Radiomics enables the extraction of the various quantitative features from the CT images for the statistical or machine learning analysis of tumor genotype. This systematic review sheds light on the CT-based radiomics model for EGFR mutation prediction in NSCLC and shows its potential for the non-invasive EGFR assessment. Methods: We conducted a very detailed literature search on the various databases, including PubMed, PubMed Central (PMC), MEDLINE, Cochrane Library, and Wiley Online Library, using combinations of keywords such as lung cancer, EGFR mutation, and radiomics. Following that, duplicate articles were removed and screened based on title, abstract, and full-text screening. Following that, a total of 14 articles was finalized. All the studies were retrospective studies evaluating the CT radiomics approach for EGFR mutation in lung cancer. Quality analysis of the studies was assessed using the Methodological Radiomics Score (METRICS), and studies were classified as moderate to excellent quality. Results: This systematic review included 14 studies that evaluated CT-based radiomics and prediction AI models for the non-invasive prediction of EGFR mutation status in NSCLC. All the studies were retrospective, and most of them were single-institution, with sample sizes ranging from approximately 80 to 700 patients. Radiomics models were consistent with the discriminatory performance, with AUC values typically between 0.70 and 0.90. Models that incorporated peritumoral features or habitat-based features outperformed those using intratumoral features alone. Integration of the radiomics with clinical variables or deep learning-derived features yielded improvement in the performance, often reporting higher AUC (~ 0.92). The above performance findings were primarily based on the internal validation. External validation was not common, and if performed, it was associated with reduced performance. Conclusions: The findings indicated that CT radiomics detects CT imaging features that correlate with EGFR-driven behavior of the tumor in NSCLC and serve as a non-invasive test. However, heterogeneity in imaging protocols, feature extraction, and modeling strategies limits generalizability. While CT radiomics cannot replace tissue biopsy or liquid biopsy in current times, CT radiomics can be utilized as a complementary tool when molecular testing is not available in resource-limited settings. Prospective, multicenter studies with standardized workflows and external validation are essential before routine clinical implementation.