Patch-Based Lightweight 3D CNNs for Anatomical Brain Segmentation Under Severe Data Scarcity: Automated DLPFC Segmentation in Structural MRI

The FDA has approved repetitive Transcranial Magnetic Stimulation (rTMS) as a treatment for major depressive disorder that affects the dorsolateral prefrontal cortex (DLPFC). However, clinical effectiveness can often be limited by inaccuracies in target localisation using heuristic methods derived from head surface analysis. Even though manual segmentation of MRI data is accurate, it takes a long time. This article discusses a three-dimensional deep learning pipeline that can segment the DLPFC quickly and without wasting resources. We have developed a custom lightweight 3D U-Net architecture that we trained from scratch with just a few data samples (N=4). The proposed method attained an average Dice similarity coefficient of 0.53 (with a peak value of 0.64) through a patch-based learning strategy and leave-one-subject-out validation that excluded one subject, while decreasing the segmentation time from 45 minutes to under 40 seconds. The findings indicate the feasibility of implementation on consumer devices for neural navigation.