# Deep learning on brain metastasis for predicting EGFR genotype and EGFR-TKI therapy response in metastatic NSCLC: a multicenter study

## TL;DR

A deep learning system was developed to predict EGFR mutations and treatment response in brain metastatic lung cancer patients using brain MRI scans.

## Contribution

A novel deep learning system (ETS) was developed to non-invasively predict EGFR mutation status and EGFR-TKI therapy response in brain metastatic NSCLC patients.

## Key findings

- The ETS achieved AUCs of 0.842, 0.833, and 0.832 for predicting EGFR mutation status across three validation cohorts.
- The fusion model combining MRI and clinical factors achieved AUCs of 0.747, 0.726, and 0.728 for predicting EGFR-TKI therapy response.
- The ETS shows potential as a non-invasive tool to guide personalized treatment decisions for metastatic NSCLC patients.

## Abstract

Brain metastases are common in patients with advanced non-small cell lung cancer (NSCLC), particularly those harboring EGFR mutations, and accurate prediction of EGFR mutation status and therapeutic response is crucial for guiding targeted therapy. This study aims to conduct a deep learning (DL) approach to automatically predict epidermal growth factor receptor (EGFR) genotype and response to EGFR-tyrosine kinase inhibitor (TKI) therapy in NSCLC patients with brain metastatic tumor (BM).

For training and validating the DL models, 388 patients were enrolled from three centers between Jul. 2014 and Dec.2022 (230 from center 1, 80 from center 2 and 78 from center 3). Contrast-enhanced T1-weighted (T1CE) and T2-weighted (T2W) brain MRI images before treatment for each patient were obtained for analyses. We developed an EGFR-TKI system (ETS) for automated detection of brain metastatic (BM) lesions and to differentiate EGFR mutation status and predict response to EGFR-TKI therapy. The models underwent rigorous evaluation through receiver operating characteristic (ROC) curve analyses, where metrics such as area under the curve (AUC), sensitivity, and specificity were examined.

For prediction of EGFR mutation status, the ETS integrating radiological-based features and clinical factors achieved AUCs of 0.842, 0.833 and 0.832 on the internal validation, external validation 1 and external validation 2 cohort, respectively. For forecasting response to EGFR-TKI therapy, the fusion model created by amalgamating MRI with clinical factors generated AUCs of 0.747, 0.726 and 0.728 on the internal validation, external validation 1, and external validation 2 cohort, respectively.

The ETS may have the potential to work as a non-invasive tool for predicting EGFR mutation status and response to EGFR-TKI therapy, which holds promise as a non-invasive tool to assist clinicians in making decisions about personalized treatment strategies.