An interpretable deep learning model for hallux valgus prediction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Hallux valgus (HV) diagnosis relies on manual measurements of hallux valgus angle (HVA) and intermetatarsal angle (IMA), which are time-consuming and prone to errors.
• Automated methods are needed to improve the efficiency and accuracy of HV diagnosis.

Purpose of the Study:

• To develop an interpretable deep learning model for automatic annotation of 12 foot landmarks.
• To automatically calculate HVA and IMA for hallux valgus diagnosis.
• To enhance the efficiency and accuracy of hallux valgus diagnosis.

Main Methods:

• A deep learning model (SE-DNN) was trained on 2,000 manually labeled foot X-ray images.
• The AG-UNet architecture was used for segmentation of key foot structures (PH1, MT1, MT2).
• Model performance was evaluated by comparing its landmark identification and angle calculations against manual measurements by surgical specialists.

Main Results:

• The model achieved an average landmark error distance between 1.9 mm and 5.6 mm, with an overall average error less than 3.1 mm.
• Inter-rater agreement for HVA and IMA measurements between the model and experts was high, with Intraclass Correlation Coefficient (ICC) results ≥ 0.9.
• The interpretable deep learning model demonstrated reliability and accuracy comparable to or exceeding manual expert measurements.

Conclusions:

• An interpretable deep learning model was successfully developed for automatic hallux valgus diagnosis.
• The model accurately identifies 12 landmarks and calculates HVA and IMA, offering significant advantages in reliability and accuracy over manual methods.
• The developed method has been successfully implemented in hospitals, demonstrating significant detection results.