Deep Learning for Automated Ischemic Stroke Lesion Segmentation from Multi-Spectral MRI

Stroke is one of the most prevalent diseases that cause long-term disability and mortality worldwide. Precisely detecting stroke lesions is crucial to diagnosing disease and planning potential treatments. Applications that assist specialists in automated lesion detection can play an important role in preventing time-consuming tasks. For quantitatively detecting strokes, specialists frequently use magnetic resonance imaging (MRI). In light of these considerations, we present a five-layer modified recurrent U-net model designed for the automated segmentation of ischemic stroke lesions in multi-spectral MRIs. The methodology implemented includes individually trained case MRI slices using the leave-one-out cross-validation (LOOCV) approach. The effectiveness of the developed model was evaluated by subject-wise metrics in comparison with the ground truth, yielding a very competitive average dice score (DSC) of 0.748.