Enhancing Lesion Segmentation via Medical Image-Mask Pair Synthesis using Phenotype-Conditioned Diffusion Models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Data scarcity and sample imbalance hinder the development of robust lesion segmentation models.
• Existing synthetic data augmentation methods often fail to generate high-quality lesion-containing samples.

Purpose of the Study:

• To introduce LesionLab, a novel framework for synthesizing medical image-mask pairs.
• To augment existing datasets, creating more balanced and diverse training data for lesion segmentation.
• To improve the quality and controllability of synthetic medical image data generation.

Main Methods:

• Phenotype-guided text prompts are designed by clustering radiomic features to capture complex lesion characteristics.
• A dual-check quality control mechanism using foundation model priors assesses sample quality and hardness.
• Synthetic data generation is controlled to precisely transform lesion-free to lesion-containing samples.

Main Results:

• LesionLab effectively augments datasets, leading to more balanced and diverse training data.
• The phenotype-guided prompts enhance the controllability of synthetic data generation.
• The dual-check mechanism filters low-quality samples and prioritizes challenging cases, improving model training.
• Experiments on three public datasets show LesionLab outperforms existing synthetic data augmentation methods.

Conclusions:

• LesionLab provides a superior approach to synthetic data augmentation for medical lesion segmentation.
• The framework addresses key challenges of data scarcity and sample imbalance effectively.
• LesionLab contributes to the development of more robust and accurate medical image segmentation models.