Modified UNet Model for Brain Stroke Lesion Segmentation on Computed Tomography Images

The task of segmentation brain regions affected by ischemic stroke is help to tackle important challenges modern imaging analysis. Unfortunately, at the moment, models for solving this problem using machine learning methods are far from ideal. In paper, we consider a modified 3D UNet architecture improve quality based on computed tomography images. We use ISLES 2018 (Ischemic Stroke Lesion Segmentation Challenge 2018) open dataset train and test proposed model. Interpretation obtained results, as well ideas further experiments included in paper. Our evaluation performed Dice or f1 score coefficient Jaccard index. may simply be extended ischemia image identification selecting relevant hyperparameters. Dice/f1 similarity our model shown 58% results close ground truth which higher than standard model, demonstrating that can accurately segment stroke. us uses an efficient averaging method inside neural network. Since set limited number, data augmentation network regularization prevent overfitting gave best result. addition, one advantages Intersection over Union loss function, assessment coincidence shapes recognized zones.