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

Accurate lesion segmentation in medical images is critical for early diagnosis and treatment planning, yet it remains challenging because training robust lesion segmentation models suffers from the data scarcity and sample imbalance issues. Synthetic data augmentation is a promising approach to address these challenges, but existing methods may struggle to produce high-quality lesion-containing samples. In this paper, we present Le sionLab, a novel framework for medical image-mask pair synthesis which augments the existing training dataset to a more balanced and diverse one. Specifically, we first design a phenotype-guided text prompt that captures complex lesion characteristics by clustering radiomic features into distinct phenotypes, improving the controllability of synthetic data generation to achieve more precise trans formation from lesion-free to lesion-containing samples. Additionally, we introduce a dual-check quality control mechanism that integrates priors of foundation models to compute quality and hardness scores, effectively filtering low-quality samples while prioritizing challenging cases for model training. Extensive experiments on three public datasets demonstrate that our proposed Lesion Lab is superior to existing synthetic data augmentation methods for lesion segmentation.