Brain hemorrhage segmentation in ct scan images using deep learning based approach

Abstract

Cerebral hemorrhage is a major cause of brain death with high early mortality and severe sequelae. Subarachnoid hemorrhage, a type of cerebral hemorrhage with small area measures and irregular shapes, is particularly difficult for image segmentation; therefore, rapid and accurate judgments of CT images of cerebral hemorrhage are vital to reduce mortality and sequelae. This research aims to propose an improved active learning method based on a convolutional neural network for small lesion cerebral hemorrhage segmentation and improve performance by iterative retraining based on segmentation output on large datasets without ground truth. After using a dataset annotated by experts to obtain the initial network model, another unannotated dataset was segmented. The data was divided into two categories: large and small area shading. The experimental process consisted of two parts: 1) a model comparison experiment and 2) an active learning comparison, where variables and pre-processing were kept consistent. Three segmentation models—U-Net, CE-Net, and FCN—were evaluated. Each model was trained three times to identify the one with the best effect. After obtaining the best model, the ground truth images of unlabeled data were obtained through improved active learning to classify high-quality and low-quality images manually in the active learning loop. The segmentation results of high-quality images were used as labels to segment low-quality images. Then, all obtained labels were merged according to a certain proportion. The segmentation results for different numbers of images were compared to determine the impact of active learning on accuracy. The results showed that, compared with original methods, the segmentation accuracy for small lesions was greatly improved. While the improved active learning proposed in this paper performed best when the dataset contained 100 to 200 images, accuracy began to decline when the dataset reached 500 images, though it still maintained an advantage. IMPLICATION OF THE THESIS: This improved segmentation method reduces the manpower needed for labeling and eliminates many professional requirements, especially in complex situations. This new active learning method not only provides a more precise diagnosis but also reduces the cost of medical treatment.