Classifying brain metastases originating from different pathological subtypes of lung cancer via a multimodal magnetic resonance imaging-based deep learning approach.

Background: The prompt, efficient, and noninvasive classification of brain metastases (BMs) originating from different pathological subtypes of lung cancer remains a challenging task to undertake in clinic. Therefore, we aimed to investigate the feasibility of a deep learning (DL) approach based on multimodal magnetic resonance imaging (MRI) in classifying BMs originating from different pathological subtypes of lung cancer. Methods: This retrospective analysis analyzed 262 patients with lung cancer BMs between August 2019 and September 2021, including 154 cases of lung adenocarcinoma (LUAD), 48 cases of small-cell lung cancer (SCLC), and 60 cases of other pathological subtypes of lung cancer (OPTLC). Multimodal MRI included T2 fluid-attenuated inversion recovery (FLAIR), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) mapping, and T1-weighted contrast enhancement (T1CE) sequences. ITK-SNAP was used to perform image segmentation in a semiautomatic manner. The largest slice of the tumor was selected on each of the four sequences, and the region of interest was drawn along the tumor edge. BM lesions with diameter greater than 1cm were drawn. The data obtained from the four sequences were randomly divided into training, validation, and testing sets at a ratio of 7:1:2. We employed the ResNet-18 DL approach as the basic classification framework and performed classification detection on T2 FLAIR, DWI, ADC, and T1CE sequences. The discrimination performances of T2 FLAIR, DWI, ADC, and T1CE sequences were assessed via receiver operating characteristic (ROC) curve analysis. Results: A total of 262 patients comprising 1,344 samples and 357 BM lesions were enrolled in the study. In the ROC curve analysis, the area under the curve (AUC) of the DL approach in classifying BMs originating from LUAD, SCLC, and OPTLC as well as the microaverage were respectively 0.71, 0.66, 0.66, and 0.74 with the T2 FLAIR sequence; 0.67, 0.65, 0.65, and 0.71 with the DWI sequence; 0.75, 0.92, 0.88, and 0.83 with the ADC sequence; and 0.74, 0.88, 0.82, and 0.83, with the T1CE sequence. Conclusions: The DL approach based on multimodal MRI was able to classify BMs originating from different pathological subtypes of lung cancer, particularly utilizing ADC and T1CE sequences. These findings provide a basis for further development of non-invasive diagnostic tools.