Deep learning for diagnosing osteonecrosis of the femoral head based on magnetic resonance imaging

Abstract

Background and objective: Early-stage osteonecrosis of the femoral head (ONFH) can be difficult to detect because of a lack of symptoms. Magnetic resonance imaging (MRI) is sufficiently sensitive to detect ONFH; however, the diagnosis of ONFH requires experience and is time consuming. We developed a fully automatic deep learning model for detecting early-stage ONFH lesions on MRI.

Methods: This was a single-center retrospective study. Between January 2016 and December 2019, 298 patients underwent MRI and were diagnosed with ONFH. Of these patients, 110 with early-stage ONFH were included. Using a 7:3 ratio, we randomly divided them into training and testing datasets. All 3640 segments were delineated as the ground truth definition. The diagnostic performance of our model was analyzed using the receiver operating characteristic curve with the area under the receiver operating characteristic curve (AUC) and Hausdorff distance (HD). Differences in the area between the prediction and ground truth definition were assessed using the Pearson correlation and Bland-Altman plot.

Results: Our model's AUC was 0.97 with a mean sensitivity of 0.95 (0.95, 0.96) and specificity of 0.97 (0.96, 0.97). Our model's prediction had similar results with the ground truth definition with an average HD of 1.491 and correlation coefficient (r) of 0.84. The bias of the Bland-Altman analyses was 1.4 px (-117.7-120.5 px).

Conclusions: Our model could detect early-stage ONFH lesions in less time than the experts. However, future multicenter studies with larger data are required to further verify and improve our model.

Keywords: Deep learning; Diagnosis; ONFH; Osteonecrosis; Osteonecrosis of the femoral head.