Hippocampus Segmentation using Patch-based Representation and ROC Label Enhancement

Magnetic Resonance Imaging (MRI) is a quantitative neuroimaging technique to support anatomical structure segmentation. Still, proper segmentation requires modeling small MRI regions, not to mention the class imbalance issue that yields false-positive predictions. This work introduces a Patch-based Segmentation with a Label Enhancement ap-proach, termed PSLE, for improved MRI-based hippocampus segmentation, by combining different texture filters to capture salient patch relationships. First, we select target-related regions to perform a convex candidate combination for label estimation. Next, we improve the overall performance by fitting the decision threshold based on the Receiver Operating Characteristic (ROC) curve, tackling the class imbalance problem. In the middle, we analyze the effect on the performance metrics of the primary hyperparameters and stages (ablation study). Finally, the state-of-the-art methods are compared with multi -atlas segmentation and deep learning algorithms in three well-known hippocampus segmentation benchmark MRI collections: LONI, ADNI, and SATA.