Segmentation of Chronic Subdural Hematomas Using 3D Convolutional Neural Networks

Abstract

Chronic subdural hematomas (cSDHs) are an increasingly prevalent neurologic disease that often requires surgical intervention to alleviate compression of the brain. Management of cSDHs relies heavily on computed tomography (CT) imaging, and serial imaging is frequently obtained to help direct management. The volume of hematoma provides critical information in guiding therapy and evaluating new methods of management. We set out to develop an automated program to compute the volume of hematoma on CT scans for both pre- and postoperative images. A total of 21,710 images (128 CT scans) were manually segmented and used to train a convolutional neural network to automatically segment cSDHs. We included both pre- and postoperative coronal head CTs from patients undergoing surgical management of cSDHs. Our best model achieved a DICE score of 0.8351 on the testing dataset, and an average DICE score of 0.806 ± 0.06 on the validation set. This model was trained on the full dataset with reduced volumes, a network depth of 4, and postactivation residual blocks within the context modules of the encoder pathway. Patch trained models did not perform as well and decreasing the network depth from 5 to 4 did not appear to significantly improve performance. We successfully trained a convolutional neural network on a dataset of pre- and postoperative head CTs containing cSDH. This tool could assist with automated, accurate measurements for evaluating treatment efficacy.