Surface spherical encoding and contrastive learning for virtual bone shape aging.

Area of Science:

• Biomedical Engineering
• Radiology
• Artificial Intelligence

Background:

• Bone shape changes are key indicators for osteoarthritis (OA) onset and progression.
• Accurate prediction of these changes can improve patient management and treatment strategies.

Purpose of the Study:

• To develop and validate a deep learning pipeline for predicting longitudinal bone shape changes in the femur up to four years in advance.
• To assess the utility of predicted bone shapes for diagnosing knee osteoarthritis (OA).

Main Methods:

• Utilized knee MRI data from the OA initiative study.
• Extracted femur bone surfaces, encoded them as shape maps using spherical coordinates.
• Trained a fully convolutional network on consecutive shape maps to predict future bone surfaces.
• Developed a novel multi-term loss function incorporating contrastive learning for training.

Main Results:

• The model accurately predicted bone shape changes with an L1 error of 0.7mm, comparable to MRI slice thickness.
• An ablation study showed that incorporating a contrastive learning term significantly improved the OA classifier's sensitivity from 0.537 to 0.709.
• The improved sensitivity approached the performance on ground truth data (0.740).

Conclusions:

• The deep learning pipeline offers a promising tool for predicting patient-specific knee osteoarthritis (OA) trajectories.
• This approach can enhance early diagnosis and personalized management of OA.
• The integration of contrastive learning in the loss function is crucial for improving predictive accuracy and diagnostic sensitivity.