Multimodal radiomics-based methods using deep learning for prediction of brain metastasis in non-small cell lung cancer with 18F-FDG PET/CT images.

Abstract

Objective: Approximately 57% of non-small cell lung cancer (NSCLC) patients face a 20% risk of brain metastases (BMs). The delivery of drugs to the central nervous system is challenging because of the blood-brain barrier, leading to a relatively poor prognosis for patients with BMs. Therefore, early detection and treatment of BMs are highly important for improving patient prognosis. This study aimed to investigate the feasibility of a multimodal radiomics-based method using 3D neural networks trained on 18F-FDG PET/CT images to predict BMs in NSCLC patients.

Approach: We included 226 NSCLC patients who underwent 18F-FDG PET/CT scans of areas, including the lung and brain, prior to EGFR-TKI therapy. Moreover, clinical data (age, sex, stage, etc.) were collected and analyzed. Shallow lung features and deep lung-brain features were extracted using PyRadiomics and 3D neural networks, respectively. A support vector machine (SVM) was used to predict BMs. The receiver operating characteristic (ROC) curve and F1 score were used to assess BM prediction performance.

Main result: The combination of shallow lung and shallow-deep lung-brain features demonstrated superior predictive performance (AUC=0.96±0.01). Shallow-deep lung-brain features exhibited strong significance (P<0.001) and potential predictive performance (coefficient>0.8). Moreover, BM prediction by age was significant (P<0.05).

Significance: Our approach enables the quantitative assessment of medical images and a deeper understanding of both superficial and deep tumor characteristics. This noninvasive method has the potential to identify BM-related features with statistical significance, thereby aiding in the development of targeted treatment plans for NSCLC patients.