A multimodal deep learning model for classifying early- versus late-stage osteonecrosis of the femoral head on biparametric MRI.

@article{li2026multimodal,
  title = {{A multimodal deep learning model for classifying early- versus late-stage osteonecrosis of the femoral head on biparametric MRI.}},
  author = {Siyuan Li and Huimin Huang and Yetong He and Xiaowan Zhang and Xiaoxia Li and Lin Wang and Wei He and Xiaorui Han and Xudong Huang and Gengxin Chen},
  journal = {BMC musculoskeletal disorders},
  year = {2026},
  month = oct,
  doi = {10.1186/s12891-026-10544-2},
  url = {https://europepmc.org/article/MED/42958225},
  note = {CC BY-NC-ND},
}

TY  - JOUR
TI  - A multimodal deep learning model for classifying early- versus late-stage osteonecrosis of the femoral head on biparametric MRI.
AU  - Siyuan Li
AU  - Huimin Huang
AU  - Yetong He
AU  - Xiaowan Zhang
AU  - Xiaoxia Li
AU  - Lin Wang
AU  - Wei He
AU  - Xiaorui Han
AU  - Xudong Huang
AU  - Gengxin Chen
PY  - 2026
DA  - 2026/10/13
JO  - BMC musculoskeletal disorders
DO  - 10.1186/s12891-026-10544-2
UR  - https://europepmc.org/article/MED/42958225
AB  - Objective  Precise early- versus late-stage stratification of osteonecrosis of the femoral head (ONFH) is critical for hip-preserving treatment selection, yet conventional MRI interpretation suffers from interobserver variability and limited objectivity. This study aimed to develop a multimodal deep learning model integrating biparametric MRI (STIR and T1-FSE) with clinical parameters (including age, sex, affected side (left, right, or bilateral), and etiology) to differentiate early- from late-stage ONFH and provide objective imaging evidence.Methods  This retrospective study included 530 patients with ONFH, yielding imaging data from 851 hips. The collected data comprised STIR sequences, T1-FSE sequences, and clinical information, including age, sex, and etiology. The dataset was randomly divided at the patient level into a training set (424 patients, 675 hips) and a test set (106 patients, 176 hips) at an 8:2 ratio. Seven models were constructed: a clinical model, a STIR model, a T1-FSE model, a bpMRI model, a STIR-clinical model, a T1-FSE-clinical model, and a bpMRI-clinical model. Early- and late-fusion strategies were used for multimodal integration. The diagnostic performance of the seven models for classifying early- versus late-stage ONFH was systematically compared, with particular emphasis on their ability to identify borderline cases at the transition between clinical stages. Model interpretability was assessed using Gradient-weighted Class Activation Mapping (Grad-CAM).Results  The early-fusion bpMRI-clinical model achieved the highest AUC of 0.944 (95% CI 0.905-0.973), accuracy 90.9%, sensitivity 89.0%, and specificity 92.9%, significantly outperforming single-sequence models (all P < 0.05). The confusion matrix showed that the model misclassified 10 late-stage cases as early-stage and 6 early-stage cases as late-stage. Grad-CAM visualizations indicated that the model primarily focused on clinically relevant regions, including the necrotic lesion boundary, areas of bone marrow edema, and subchondral structural abnormalities. These attention patterns were highly consistent with the core imaging features used for early- versus late-stage ONFH stratification.Conclusion  This study presents an interpretable multimodal deep learning framework that integrates dual-sequence MRI with clinical parameters. The framework accurately differentiates early- from late-stage ONFH according to the ARCO classification and provides objective evidence for determining the critical threshold for hip-preservation decision-making.Trial registration  The trial was registered with the Chinese Clinical Trial Registry (Registration No: ChiCTR2300075354. Sep 1, 2023).
SN  - 1471-2474
ER  -