Leveraging deep learning for nonlinear shape representation in anatomically parameterized statistical shape models.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• Statistical shape models (SSMs) are crucial for anatomical structure assessment.
• A limitation of current SSMs is the lack of clear links between shape coefficients and clinical parameters.

Purpose of the Study:

• To propose a novel deep learning-based anatomically parameterized SSM (DL-ANAT_SSM).
• To establish a nonlinear relationship between anatomical parameters and bone shape information for improved model interpretability and precision.

Main Methods:

• Utilized a multilayer perceptron model trained on synthetic and real femoral bone datasets.
• Learned nonlinear mapping between anatomical measurements and shape parameters.
• Compared DL-ANAT_SSM performance against a linear SSM baseline.

Main Results:

• DL-ANAT_SSM showed superior performance in predicting 3D bone shape from anatomical parameters on unseen data.
• Fine-tuning the model further improved its predictive performance.
• Demonstrated a more precise and interpretable SSM controlled by clinical parameters.

Conclusions:

• DL-ANAT_SSM offers a more precise and interpretable approach to SSMs.
• The method holds promise for morphometry analysis and patient-specific 3D model generation without preoperative imaging.