Reshapeit: reliable shape interaction with implicit template for medical anatomy reconstruction.

Area of Science:

• Medical imaging
• Computer vision
• Computational anatomy

Background:

• Accurate shape modeling of volumetric medical images is vital for quantitative analysis and surgical planning.
• Current deep learning methods for automatic shape reconstruction face limitations in image resolution and shape prior constraints.

Purpose of the Study:

• To develop a novel method for reliable and accurate anatomical shape modeling in continuous space.
• To overcome limitations of existing deep learning models in medical image shape reconstruction.

Main Methods:

• Introducing the Reliable Shape Interaction with Implicit Template (ReShapeIT) network.
• Representing anatomical structures using continuous implicit fields instead of discrete voxel grids.
• Utilizing a category-specific implicit template field combined with a deformation network and a Template Interaction Module (TIM) for shape encoding and refinement.

Main Results:

• ReShapeIT was evaluated on Liver, Pancreas, and Lung Lobe datasets.
• The method demonstrated superior performance in 3D shape reconstruction compared to state-of-the-art approaches.
• Achieved competitive Chamfer Distance/Earth Mover's Distance scores across all tested anatomical structures.

Conclusions:

• ReShapeIT offers a reliable and generalizable approach to implicit anatomical shape modeling.
• The method effectively leverages shared template priors and instance-level deformations.
• The code for ReShapeIT is publicly available for further research and application.