Local Style Transfer via Latent Space Manipulation for Cross-Disease Lesion Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for automatic lesion segmentation requires large datasets, which are often unavailable in clinical settings.
• Leveraging external datasets is a viable strategy to overcome data scarcity in medical image analysis.
• Existing methods struggle with cross-disease lesion segmentation, especially in low-data regimes.

Purpose of the Study:

• To introduce LatenTrans, a novel framework for boosting lesion segmentation performance in extremely low data conditions.
• To enable effective knowledge transfer from external datasets for improved segmentation accuracy.
• To address the challenge of utilizing diverse datasets for training segmentation models.

Main Methods:

• Projecting images into a latent space using aligned style-based generative models to encode lesion semantics.
• Employing a consistency-aware latent code manipulation module for local style transfer from non-target to target-like lesions.
• Proposing a new metric, Normalized Latent Distance, for selecting optimal external datasets for knowledge transfer.

Main Results:

• LatenTrans successfully translates non-target lesions into target-like representations, expanding the training dataset.
• The framework demonstrates superior performance in cross-disease lesion segmentation compared to existing methods.
• Experiments on lung and brain lesion segmentation validate the effectiveness of LatenTrans in low-data regimes.

Conclusions:

• LatenTrans offers a powerful solution for improving lesion segmentation accuracy when training data is limited.
• The proposed method facilitates effective knowledge transfer across different lesion types and datasets.
• This framework has significant potential for enhancing diagnostic assistance in clinical practice.