Development and evaluation of a 3D ensemble framework for automatic diagnosis of early osteonecrosis of the femoral head based on MRI: a multicenter diagnostic study

Abstract

Background: Efficient and reliable diagnosis of early osteonecrosis of the femoral head (ONFH) based on MRI is crucial for the formulation of clinical treatment plans. This study aimed to apply artificial intelligence (AI) to achieve automatic diagnosis and visualization of early ONFH, thereby improving the success rate of hip-preserving treatments.

Method: This retrospective study constructed a multicenter dataset using MRI data of 381 femoral heads from 209 patients with ONFH collected from four institutions (including 239 early ONFH cases and 142 non-ONFH cases). The dataset was divided into training, validation, and internal and external test datasets. This study developed a 3D ensemble framework to automatically diagnose early osteonecrosis of the femoral head based on MRI and utilized 3D Grad-CAM to visualize its decision-making process. Finally, the diagnostic performance of the framework was experimentally evaluated on the MRI dataset and compared with the diagnostic results of three orthopedic surgeons.

Results: On the internal test dataset, the 3D-ONFHNet framework achieved overall diagnostic performance with an accuracy of 93.83%, sensitivity of 89.44%, specificity of 95.56%, F1-score of 87.67%, and AUC of 95.41%. On the two external test datasets, the framework achieved overall diagnostic accuracies of 87.76% and 87.60%, respectively. Compared to three orthopedic surgeons, the diagnostic performance of the 3D-ONFHNet framework was comparable to that of senior orthopedic surgeons and superior to that of junior orthopedic surgeons.

Conclusions: The framework proposed in this study can generate staging results for early ONFH and provide visualizations of internal signal changes within the femoral head. It assists orthopedic surgeons in screening for early ONFH on MRI in a clinical setting, facilitating preoperative planning and subsequent treatment strategies. This framework not only enhances diagnostic efficiency but also offers valuable diagnostic references for physicians.

Keywords: MRI; artificial intelligence; clinical decision-making; osteonecrosis of the femoral head; predictive model.

Figure 1

Multicenter data composition diagram.

Figure 2

The overall 3D ensemble framework proposed in this study, 3D-ONFHNet. 3D, three-dimensional; ONFH, osteonecrosis of the femoral head.

Figure 3

Architecture of the 3D transition multiPool module in the staging subnet. 3D, three-dimensional.

Figure 4

Overview of the experimental workflow. (a) Data preprocessing: the collected raw data are preprocessed to ensure consistent data quality. (b) Model training: the preprocessed data are then provided as input to the deep learning model, and the model parameters are optimized to achieve the best possible performance. (c) Internal testing: the trained model is evaluated using internal test datasets to assess its performance. (d) External testing: external datasets are employed to validate the model's generalizability and robustness.

Figure 5

ROC curve (A) and confusion matrix (B) for evaluating the staging performance of the 3D-ONFHNet framework in the internal test dataset. ROC, receiver operating characteristic.

Figure 6

ROC curve (A,C) and confusion matrix (B,D) for evaluating the staging performance of the 3D-ONFHNet framework in the external test dataset. ROC, receiver operating characteristic.

Figure 7

Visualization of 3D heatmaps for stage I-II ONFH and 3D reconstructed heatmap images. Key slices of the femoral head are selected for display in the 3D heatmap visualization. 3D, three-dimensional; ONFH, osteonecrosis of the femoral head.