Automatic Segmentation of Intracranial Hemorrhage in Computed Tomography Scans with Convolution Neural Networks

Purpose: Intracranial hemorrhage (ICH) is a serious health problem requiring prompt and intensive medical treatment. The delineation of hemorrhage areas and the estimation of hemorrhage volume play a crucial role for subsequent treatment planning. In clinical practice, physicians need to inspect computed tomography (CT) scans carefully and delineate the hemorrhage area slice-by-slice, which is time-consuming and error-prone. Hence, we propose a segmentation framework based on deep learning to assist physicians in automatically delineating the area of intracranial hemorrhage and evaluating the volume of hemorrhage in CT scans. Methods: The framework is a two-staged encoder-decoder convolution neural network, which segments the coarse hemorrhage area in the first stage and the fine hemorrhage area in the second stage. With the segmentation result, the volume of hemorrhage is further estimated. The proposed framework was trained on a training set with 280 CT scans and tested on a blind test set with 121 CT scans. Results: Our model achieves a dice coefficient of 84.54%, which outperforms the single-stage segmentation model. As for hemorrhage volume estimation, our model is far ahead of the commonly used Tada formula in clinical practice and is able to obtain the more robust prediction for different hemorrhage situations. Conclusion: This cascade framework assists in successfully segmenting intracranial hemorrhage areas and evaluating the volume of hemorrhage in CT scans and can be used in clinical practice, which has the potential to be applied in medical practice.