Progressive DeepSSM: Training Methodology for Image-To-Shape Deep Models.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Deep learning in healthcare

Background:

• Statistical shape modeling (SSM) is crucial for quantitative analysis of anatomical shapes in medicine.
• Directly using 3D medical images for SSM construction requires extensive preprocessing.
• Current deep learning methods for direct SSM construction from images show suboptimal performance.

Purpose of the Study:

• To propose a new training strategy, progressive DeepSSM, to improve deep learning models for image-to-shape analysis.
• To enable learning of both coarse and fine anatomical shape features progressively.
• To enhance the accuracy and stability of SSMs generated directly from unsegmented medical images.

Main Methods:

• A multiscale training strategy named progressive DeepSSM is introduced.
• The model training progresses through multiple scales, with each scale building upon the previous one.
• Segmentation-guided multi-task learning and deep supervision loss are employed to incorporate shape priors and ensure learning at each scale.

Main Results:

• Models trained with the progressive DeepSSM strategy demonstrate superior quantitative and qualitative performance.
• The proposed strategy significantly improves the accuracy of statistical shape representations.
• Enhanced training stability is observed for deep learning models using this methodology.

Conclusions:

• Progressive DeepSSM is an effective training methodology for deep learning-based image-to-shape models.
• This strategy can be broadly adopted to improve existing deep learning methods for anatomical shape analysis.
• The approach facilitates more accurate and stable inference of statistical shape models directly from medical images.