Improving Stroke Segmentation and Classification Performance Using a Goal-Oriented Deep Learning Framework

Abstract

CT scans play a crucial role in diagnosing and planning treatment for strokes, offering essential insights into the location, size, and extent of bleeding in brain tissue. This study explores two distinct scenarios for stroke detection, classification, and segmentation, utilizing 6951 brain CT images from the TEKNOFEST competition. In both scenarios, CT images undergo preprocessing steps involving skull-stripping, normalization, and image augmentation. In the first scenario, stroke presence-absence classification achieved a 98% success rate on test images. Subsequent segmentation of images with strokes resulted in Dice scores of 59% for ischemic stroke and 67% for hemorrhagic stroke on test images. The classification of stroke types as ischemic and hemorrhagic achieved a 100% success rate, with a 97% success rate when directly classifying stroke types in images without segmentation. This indicates a 3% performance improvement when applying the classification process after stroke region segmentation. In the second scenario, a three-class classification of no stroke, ischemic stroke, and hemorrhagic stroke achieved an average of 97% success on test images. Post-classification, separately created models for the segmentation of ischemic and hemorrhagic strokes yielded Dice scores of 78% and 79%, respectively. The second scenario demonstrated a performance improvement of 19% and 12% for the segmentation of ischemic and hemorrhagic strokes through the post-classification segmentation process. The proposed approach outperforms competing teams in the competition rankings.