Joint reconstruction and segmentation in undersampled 3D knee MRI combining shape knowledge and deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Task-adapted neural networks (NNs) optimize image reconstruction for tasks like segmentation but require substantial hardware.
• Current methods often use simple NN building blocks, lacking integration of model-specific knowledge.
• End-to-end trainable methods offer potential but face computational challenges.

Purpose of the Study:

• To enhance end-to-end trainable task-adapted image reconstruction by incorporating statistical shape models (SSMs).
• To apply this to the clinically relevant problem of bone and cartilage segmentation in 3D knee MRI.
• To compare the proposed method against simultaneous multitask learning (MTL) and a complex SSMs-informed segmentation pipeline (SIS).

Main Methods:

• Extended an end-to-end trainable method with SSMs for prior information and regularization.
• Utilized SSMs to regularize segmentation maps as a post-processing step.
• Compared performance against MTL and SIS approaches for reconstruction and segmentation.

Main Results:

• The combined approach of joint end-to-end training and SSM regularization significantly improved segmentation accuracy, reducing mean and maximal surface errors.
• Achieved segmentation quality comparable to the complex SIS pipeline.
• Demonstrated a five-fold reduction in model parameters and an order of magnitude speedup in computation.

Conclusions:

• Integrating SSMs into MTL for 3D knee MRI reconstruction and segmentation offers a computationally efficient yet highly accurate solution.
• The method achieves high-quality segmentation even with significant undersampling (R=8).
• This approach balances the benefits of deep learning with the prior knowledge of SSMs for improved medical image analysis.