A Comparative Study of Segmentation Algorithms for Intracerebral Hemorrhage Detection

Abstract

Segmentation in the medical field has special importance. One of the purposes of segmentation is to visualize the area affected by the disease after disease detection in any organ. In recent years, efficient studies have been carried out for this purpose with deep learning models. In this study, three segmentation algorithms were compared for the detection of hemorrhage in brain parenchyma. These algorithms are the most familiar: U-net, LinkNet, and FPN algorithms. For the background of these algorithms, five backbones consisting of deep learning models were used. These backbones are Resnet34, ResNet50, ResNet169, EfficientNetB0, and EfficientNet B1. An original dataset was created for the study. The dataset in the study was verified by experts. In the study, the Dice coefficient and Jaccard index, which are the most common metrics in the medical field, were chosen as evaluation metrics. Considering the performance results of the algorithms, the FPN architecture with a 0.9495 Dice coefficient value for the training data and LinkNet with a 0.9244 Dice coefficient for the test data gave the best results. In addition, EfficientNetB1 provided the best results among the backbones used. When the results obtained were examined, better segmentation performance was obtained than in existing studies.