Development and validation of a deep learning model for automatic detection of depressed skull fractures from CT scans.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, Faculty of Medicine, Maharaj Nakorn Chiang Mai Hospital, Chiang Mai University, Chiang Mai, Thailand.
- Global Health and Chronic Conditions Research Center, Chiang Mai University, Chiang Mai, Thailand.
- Lampang Hospital, Lampang, Thailand.
- Department of Computer Science, Faculty of Science, Chiang Mai University, Chiang Mai, Thailand.
- Department of Diagnostic and Therapeutic Radiology, Faculty of Medicine, Ramathibodi Hospital, Mahidol University, Bangkok, Thailand.
- Department of Computer Science, Faculty of Science, Chiang Mai University, Chiang Mai, Thailand. [email protected].
- Global Health and Chronic Conditions Research Center, Chiang Mai University, Chiang Mai, Thailand. [email protected].
- Data Science Research Center, Faculty of Science, Chiang Mai University, Chiang Mai, Thailand. [email protected].
Abstract
Depressed skull fractures with bone depression greater than in one cortex might cause major consequences and require surgery in traumatic head injury patients. Therefore, skull fractures with depression in more than one cortex must be identified quickly and accurately. This study proposes using a deep learning model to deal with the task. Cranial CT scans of traumatic head injury patients with and without depressed skull fractures were collected for this retrospective investigation. A real-time object detection model, You Only Look Once (YOLO), was adopted to detect depressed skull fractures in more than one cortex. We proposed a two-phase training strategy for training the model. The model was evaluated using internal and external test datasets. The detection performance was reported in terms of accuracy, sensitivity, specificity, precision, negative predictive value, F1-score, and area under the receiver operating characteristic curve. The deep learning model demonstrated strong performance on an internal test dataset (accuracy = 0.957); however, its performance declined on two external test datasets (accuracy = 0.884 and 0.857). This model enables automated detection of depressed skull fractures, streamlining the clinical workflow by flagging high-priority cases for expedited radiologist review.