Abstract
Our proposed CNN model can effectively segment the region where the femoral head is in MRI and can identify the region of early stage femoral head necrosis for the purpose of aiding diagnosis.
Introduction
Osteonecrosis of the Femoral Head (ONFH) is a progressive and debilitating condition characterized by the death of bone tissue due to impaired blood supply. If left untreated, it often leads to the collapse of the femoral head and subsequent hip joint destruction. Early diagnosis and accurate staging are critical for effective surgical intervention. Clinicians typically rely on the Steinberg classification system, which categorizes the severity into Mild, Moderate, or Severe based on the percentage of the femoral head affected by necrosis.
In clinical practice, calculating the exact percentage of necrotic tissue from MRI scans is a labor-intensive and subjective task. Most surgeons rely on visual estimates, which can lead to significant inter-observer variability and potential misdiagnosis. To address this, there is a pressing need for automated tools that can provide objective, pixel-level quantification of necrosis to assist in clinical decision-making.
Methods
The researchers developed a multi-stage framework designed to automate the segmentation and grading of early-stage ONFH. The core of this framework is a customized Convolutional Neural Network (CNN) based on the U-Net architecture.
Segmentation and Supervision
The model enhances the standard U-Net by incorporating two additional supervision paths: geometric supervision and pixel-level supervision. These constraints are applied during training to ensure the model accurately captures the boundaries and shape of the femoral head, reducing the noise typically associated with medical image segmentation.
Necrosis Extraction and Quantification
Following the segmentation of the femoral head, the system utilizes an adaptive thresholding method to isolate necrotic tissue. This algorithm identifies the optimal intensity threshold to differentiate healthy bone from necrotic regions within the segmented area. The final stage involves calculating the ratio of necrotic pixels to the total pixels of the femoral head.
Datasets and Preprocessing
The study utilized a retrospective dataset of 261 axial T1-weighted MRI slices from patients at the Second Affiliated Hospital of Harbin Medical University. The ground truth labels were established by three orthopedic surgeons and verified by a chief surgeon. Preprocessing involved normalizing and cropping the images to a standard $256 \times 256$ resolution. The dataset was partitioned into an 80% training set and a 20% testing set.
Results
The proposed model demonstrated superior performance compared to standard architectures like LinkNet and the original U-Net. Key performance metrics included:
- Segmentation Accuracy: 97.73%
- Sensitivity: 91.17%
- Specificity: 99.40%
- Overall Diagnostic Accuracy: 90.80%
Quantification & Grading
The framework successfully categorized the necrosis stages according to Steinberg criteria: Mild (< 15%), Moderate (15%–30%), and Severe (> 30%). By providing a precise pixel count, the model offers a level of diagnostic granularity that is difficult to achieve manually.
Discussion and Limitations
A key innovation of this work is the integration of geometric constraints with deep learning. By moving away from a "black box" approach and using adaptive thresholding for the final quantification, the results are more interpretable for clinicians. This hybrid approach addresses the clinical need for both accuracy and transparency, potentially reducing human error caused by diagnostic pressure and fatigue.
Despite these promising results, the study has limitations. The use of only axial MRI slices means the model does not yet account for the full 3D volume of the femoral head. Additionally, the relatively small dataset size and the lack of external validation on data from other institutions are areas for future research. The authors plan to explore 3D reconstruction to further refine diagnostic precision.
Conclusion
This CNN-based framework provides a robust and objective tool for the early diagnosis and classification of ONFH. By automating the calculation of necrotic area, it reduces the diagnostic burden on clinicians and offers a path toward more standardized and accurate surgical planning.
Summary updated