Disentangled representation learning in cardiac image analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical images contain both anatomical information and modality-specific characteristics.
• Existing methods often struggle to disentangle these distinct factors within imaging data.

Purpose of the Study:

• To develop a novel network, Spatial Decomposition Network (SDNet), for disentangling 2D medical images into spatial anatomical and non-spatial modality factors.
• To demonstrate the utility of this factorized representation for various medical image analysis tasks.

Main Methods:

• Proposed Spatial Decomposition Network (SDNet) to factorize 2D medical images into anatomical and modality-specific representations.
• Applied the model to semi-supervised segmentation, multi-task learning (segmentation and regression), and image-to-image synthesis tasks.
• Investigated latent-space arithmetic for cross-modality synthesis and modality prediction.

Main Results:

• SDNet achieved performance comparable to fully supervised models in segmentation tasks using significantly less labeled data.
• The factorized representation improved performance in multi-task learning settings and when aggregating multimodal data (e.g., MRI and CT).
• Demonstrated successful synthesis of CT images from MRI and vice versa by manipulating modality factors, and accurate modality prediction from learned factors.

Conclusions:

• SDNet effectively disentangles anatomical and modality factors from medical images, offering a powerful representation for downstream tasks.
• The proposed method enhances efficiency in medical image analysis by reducing reliance on extensive labeled datasets.
• The learned factorized representation facilitates cross-modality image synthesis and understanding of imaging characteristics.