A team of researchers from UC San Francisco's Department of Radiology and Biomedical Imaging and Center for Intelligent Imaging (ci2) tested a learning model for intracranial hemorrhage detection and segmentation.
The researchers share their conclusions in "Semisupervised Learning for Generalizable Intracranial Hemorrhage Detection and Segmentation," published in RSNA Radiology: Artificial Intelligence. Esther Yuh, the second author, is a Professor of Radiology and Biomedical Imaging and ci2 member. Emily Lin from UCSF's Department of Radiology and Biomedical Imaging also contributed to the research.
The authors performed the retrospective study using a semisupervised learning model compared with a baseline model trained on labeled data using area under the receiver operating characteristic curve (AUC), Dice similarity coefficient (DSC), and average precision (AP) metrics.
The authors trained a "teacher" deep learning model on 457 pixel-labeled head CT scans and pseudo-labels on a separate unlabeled corpus of 25,000 examinations. A second "student" model was trained on the pixel- and pseudo-labeled dataset of scans. The investigators used a dataset of 481 scans from India, CQ500, to test for classification and segmentation and to evaluate out-of-distribution generalizability.
"The semisupervised model achieved statistically significantly higher examination AUC on CQ500 compared with the baseline," the authors write. It also achieved a higher DSC and Pixel AP compared with the baseline.
"The addition of unlabeled data in a semisupervised learning framework demonstrates stronger generalizability potential for intracranial hemorrhage detection and segmentation compared with a supervised baseline," Yuh and Lin write.
Read more about news and research at UCSF ci2.
