A Unique Block-Based 3D U-Net Deep Learning Model for Brain Metastasis MRI Image Segmentation and Classification

Brain metastasis (BM) is one of the primary neurologic complications of cancer, frequently originating in patients with lung cancer, breast cancer, colorectal cancer, and melanoma. Scarcity of data has limited the progression of deep learning studies to automate the classification and segmentation of BM magnetic resonance imaging (MRI) images. Furthermore, due to the nature of BM's, the available data is highly unbalanced with relatively diminutive cancer regions. This paper presents a novel 3D U-net deep learning model on the first public BM database. This study uses a novel technique to split the BM MRI images into smaller blocks, on which machine learning is performed primarily focused on cancer regions. By combining segmentation and classification loss functions, machine learning models can perform both BM segmentation and cancer region classification. two models are presented, the Aggressive model which is tailored towards precise segmentation that reached a Dice score of 0.88; and the Passive model which is suited towards detecting blocks that contain cancer with fewer false positives. Combining both models can detect 84% of all cancer cases. When both models agree, the prediction has a high positive predictive value (84.7% for cancer; 99.3% for non-cancer) while when the models differ, the value is low (11.9% for cancer) and would merit further physician review.