Deep learning MRI model using perifocal edema to differentiate brain metastasis from central nervous system infection

Objectives This study aims to establish a deep learning (DL) model based on peripheral edema zones using multi-sequence MRI to distinguish between brain metastases (BM) and central nervous system infections (CNSI). Methods Retrospective data collection was done on 214 patients at medical institution A, and were randomly divided into a training set and an internal validation set with a ratio of 4:1. 60 patients from medical institution B were used as an external validation set. The volume of interest (VOI) was manually delineated based on T2-weighted imaging (T2WI), T1-weighted imaging (T1WI), diffusion-weighted imaging (DWI), fluid-attenuated inversion recovery imaging (FLAIR) and T1-enhanced imaging (T1C). Using the three-dimensional residual network-18 (3D ResNet-18) architecture based on perilesional edema, single or multiple sequence MRI DL models were developed. Gradient-weighted class activation mapping was used to generate heat maps to visualize the model. The area under the curve (AUC) was used to evaluate the predictive efficiency of each DL model. Results The three-sequence model combined DWI, T1C, and T2WI performed the best in the training and internal validation set, with an AUC of 0.996 and 0.879. In the external validation set, the DWI single-sequence model had an AUC of 0.798, making it the optimal DL model. The edema region adjacent to the enhanced lesion received more attention in the heatmap of T1C, whereas the entire edema area had relatively scattered focus in that of T2WI and DWI. Conclusions The 3D ResNet-18 model established by the perilesional edema could effectively distinguish between BM and CNSI.