IEEE Access (Jan 2024)

Applications of Deep Learning Models on the Medical Images of Osteonecrosis of the Femoral Head (ONFH): A Comprehensive Review

  • Jiao Wang,
  • Yi Zheng,
  • Jun Luo,
  • Timothy Tin-Yan Lee,
  • Pengfei Li,
  • Ying-Qi Zhang,
  • James Chung-Wai Cheung,
  • Duo Wai-Chi Wong,
  • Ming Ni

DOI
https://doi.org/10.1109/ACCESS.2024.3389669
Journal volume & issue
Vol. 12
pp. 57613 – 57632

Abstract

Read online

Deep learning models have demonstrated promising results in the early and accurate diagnosis of osteonecrosis of the femoral head (ONFH), enabling early detection and informed surgical decision-making. The objective of this review is to summarize the applications of deep learning models on the medical images of ONFH. English papers were searched from CINAHL via EBSCOhost, Embase, IEEE Xplore® Digital Library, PubMed, Scopus, and Web of Science. Sixteen studies (n =16) were eligible for data synthesis. Among these, five studies (n =5) focusing on radiographs, ten studies (n =10) focusing on magnetic resonance imaging, and one study (n =1) focusing on computed tomographic images. The applications of these studies included identifying ONFH from normal or other hip pathologies, classifying severity, segmenting, and detecting femoral head and necrotic regions, predicting signs and symptoms of ONFH, and predicting potential ONFH after fracture fixation. Generally, the models demonstrated good to excellent classification performance and excellent discriminatory power; and generally comparable to that of experienced physicians and superior to that of less experienced physicians. However, the external validity of these studies demonstrated only moderate, as evidenced by the performance on the external testing set and might be attributed to the relatively small data size used during model training. we observed a shift from CNN-based models to U-Net-based models (i.e., with encoder-decoder architecture). In addition to streamlining the segmentation, detection, and classification procedures, future studies will explore multimodal attention, self-supervised learning, explainable models, and data augmentation through generative models.

Keywords