A team from Shahid Beheshti University and University of Virginia reviews machine learning and deep learning radiomics models to predict EGFR mutation status in non-small cell lung cancer brain metastases, highlighting a pooled AUC of 0.91 and strong clinical potential.
Key points
- Meta-analysis of 20 studies comprising 3,517 patients and 6,205 NSCLC brain metastatic lesions.
- Radiomics-based ML (LASSO, SVM, RF) and DL (ResNet50) models analyze MRI features to predict EGFR mutation status.
- Best-performance models achieve pooled AUC of 0.91 (95% CI: 0.88–0.93) and accuracy of 0.82.
- Sensitivity is 0.87 and specificity 0.86, yielding a diagnostic odds ratio of 35.2.
- Subgroup analysis shows no significant performance difference between ML and DL approaches.
Why it matters: Noninvasive, accurate EGFR status prediction can guide timely targeted therapies and reduce the need for risky biopsies in metastatic lung cancer. These high-performance ML and DL radiomics tools could reshape personalized treatment planning and improve patient outcomes in NSCLC brain metastases.
Q&A
- What is EGFR and its role in NSCLC brain metastases?
- What are radiomics features in MRI analysis?
- How do machine learning and deep learning differ here?
- What does AUC indicate in diagnostic studies?
- What limitations affect current ML models for EGFR prediction?
