Deep learning model for differentiating acute and chronic osteoporotic vertebral compressive fractures on multidetector CT: a retrospective study with external validation
TL;DR
The YOLO v7 model achieved good performance for CT-based differentiation of acute versus chronic OVCFs in patients and showed better performance compared to radiologists in both testing sets, which may serve as a decision-support tool for CT-equivocal cases.
Abstract
Background Osteoporotic vertebral compression fractures (OVCFs) are a common spinal disease. Differentiating acute and chronic fractures is the key to determining the treatment plan. To develop and evaluate a deep learning model capable of differentiating acute and chronic OVCFs on CT images. Methods The internal dataset comprised CT images of 624 fractured vertebrae from 400 patients with OVCF treated at Hospital 1 between January 1, 2020, and November 1, 2023. The patients were randomly divided into a training set (340 patients, 532 fractured vertebrae, 85.0%) and an internal testing set (60 patients, 92 fractured vertebrae, 15.0%). An external testing set included CT images of 86 fractured vertebrae from 70 patients with OVCF treated at Hospital 2 from January 1, 2023, to November 1, 2023, using identical inclusion criteria. Three radiologists manually delineated regions of interest (ROIs) in CT images of diseased vertebrae using Anaconda Prompt software with rectangular bounding boxes. The trained YOLO v7 model’s performance was evaluated on both the internal and external testing sets and compared with that of the three radiologists using metrics with accuracy, sensitivity, specificity, and AUC. Results The internal testing set revealed the following: AUC, 0.937 (95%CI: 0.932–0.991); accuracy, 0.961 (0.930–0.977); sensitivity, 0.973 (95%CI: 0.958–0.994); specificity, 0.761 (95%CI: 0.728–0.784); precision, 0.936 (95%CI: 0.920–0.966), and F1 score; 0.954 (95%CI: 0.913–0.975). The external testing set revealed the following: AUC, 0.882 (95%CI: 0.839–0.894); accuracy, 0.852 (95%CI: 0.827–0.886); sensitivity, 0.895 (95%CI: 0.840–0.913); specificity, 0.780 (95%CI: 0.730–0.812); precision, 0.874 (95%CI: 0.839–0.892); and F1 score, 0.884 (95%CI: 0.846–0.903). Conclusion The YOLO v7 model achieved good performance for CT-based differentiation of acute versus chronic OVCFs in patients and showed better performance compared to radiologists in both testing sets, which may serve as a decision-support tool for CT-equivocal cases.