A combined deep learning method better detected hemorrhages and identified different subtypes of intracranial hemorrhage than single algorithms used alone, according to a new study published in the Journal of Digital Imaging.
“Detecting intracranial hemorrhage and segmenting using a computer-aided detection or diagnosis (CAD)-based automatic system is a promising approach in improving workflow and reducing human errors. It results in a better patient outcome,” wrote first author Junghwan Cho, of CAIDE Systems Inc. in Lowell, Massachusetts, and colleagues.
All previous studies, according to the authors, have focused on binary classification problems, but haven’t handled the five subcategories of intracranial hemorrhage. Their novel method, they noted, focuses on those categories as well as the segmentation of lesions.
The team trained a cascade deep learning model and dual fully convolutional networks (FCNs) using more than 135,000 CT images from 5,702 patients. Ten experts labeled the images to establish a ground truth, including more than 33,000 labeled as bleeding.
By combining the two, the team could achieve a better binary classification and improved segmentation of hemorrhagic lesions compared to a model utilizing one or the other. The method identified bleeding with 98 percent sensitivity—a 1 percent improvement compared to singular CNNS—and maintained a 99 percent specificity.
The combined technique achieved an 80 percent accuracy for detecting hemorrhage type and segmenting their lesions as well as 82 percent recall rate—a 3.44 percent improvement from a single FCN model.