Mask Correction in 3-D Tomography Brain Images for Weakly Supervised Segmentation of Acute Ischemic Stroke

In this paper we propose a method for weakly supervised segmentation of 3-D computed tomography brain images of acute ischemic stroke using convolutional neural nets. To improve the segmentation quality of stroke areas, the concepts of a distance map and weight map are introduced. The maps are utilized to correct the predictions of the model at the boundaries of the affected areas. Additionally, a smoothing method is introduced for segmentation masks to reduce labeling defects. The study uses two sets of data: the primary set that includes labeling made by a single radiologist, and the auxiliary set of smaller size with several variants of labeling made by different radiologists. The latter set is analyzed to reveal the basic characteristics of labeling discrepancies which arise due to complex nature of analyzed images. The 3D U-Net model is employed for the primary set segmentation. DICE loss and Focal loss are used to train the model, and DICE score is utilized to evaluate the quality of forecasts. The results of experiments demonstrate the effectiveness of the proposed method.