ShapeMed-Knee: A Dataset and Neural Shape Model Benchmark for Modeling 3D Femurs

Area of Science:

• Medical imaging and computational anatomy.
• Biomedical data science and machine learning.
• Orthopedics and osteoarthritis research.

Background:

• Accurate analysis of anatomic shapes is crucial for diagnosing diseases like osteoarthritis, affecting millions.
• Existing methods for 3D shape analysis in medicine have limitations in accuracy and clinical utility.

Purpose of the Study:

• Introduce ShapeMed-Knee, a comprehensive 3D dataset for knee osteoarthritis research.
• Develop and evaluate novel neural shape models for improved diagnostic and prognostic capabilities.
• Provide benchmarks for assessing shape reconstruction and clinical prediction tasks.

Main Methods:

• Created ShapeMed-Knee, a dataset of 9,376 high-resolution 3D knee shapes (femur and cartilage) from medical imaging.
• Developed a novel hybrid explicit-implicit neural shape model.
• Evaluated model performance on reconstruction accuracy and clinical prediction tasks, including osteoarthritis feature prediction.

Main Results:

• The hybrid neural model achieved up to 40% better reconstruction accuracy compared to existing models.
• Achieved state-of-the-art performance in preserving cartilage biomarkers (RMSE ≤ 0.05).
• First models to successfully predict localized osteoarthritis features (e.g., osteophyte size/localization) with 63% accuracy.

Conclusions:

• ShapeMed-Knee dataset and benchmarks facilitate 3D modeling applications in medicine, reducing barriers.
• Advancements in 3D shape modeling enhance osteoarthritis diagnosis and risk stratification.
• The freely accessible dataset, code, and benchmarks support further research in medical imaging analysis.