An Explainable Plane-Wise ConvNet Approach for Detecting Femoral Head Osteonecrosis from Magnetic Resonance Images

Area of Science:

• Orthopedic imaging
• Artificial intelligence in medicine
• Medical decision support

Background:

• Osteonecrosis of the femoral head (ONFH) diagnosis is challenging in early stages due to subtle radiological findings.
• Delayed or inaccurate staging of ONFH can lead to femoral head collapse and functional loss.
• Magnetic resonance imaging (MRI) is sensitive but interpretation varies with observer experience, necessitating automated solutions.

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for classifying ONFH into early (Stage I-II) and late (Stage III-IV) stages.
• To assess the performance of plane-wise (axial and coronal) MRI analysis for ONFH staging.
• To create an explainable decision support tool to reduce observer variability in ONFH staging.

Main Methods:

• A deep learning model using ConvNeXt Tiny was trained on MRI images for ONFH classification.
• Axial and coronal MR images were processed separately, including resizing, normalization, and augmentation.
• Weighted loss and optimized decision thresholds were employed to handle class imbalance and prioritize critical late-stage cases.

Main Results:

• The axial plane model achieved 94.51% accuracy and 0.981 AUC; the coronal plane model achieved 92.84% accuracy and 0.988 AUC.
• Both models demonstrated high sensitivity and specificity in distinguishing early from late-stage ONFH.
• Grad-CAM visualizations confirmed the model focused on clinically relevant femoral head regions.

Conclusions:

• Deep learning-based plane-wise MRI analysis effectively distinguishes early and late-stage ONFH with high performance.
• The proposed explainable AI approach can serve as a valuable decision support tool for clinicians.
• Future research should focus on external validation with multicenter datasets and paired patient-level images.