Fair Text to Medical Image Diffusion Model with Subgroup Distribution Aligned Tuning.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Text-to-Medical Image (T2MedI) synthesis using latent diffusion models can address medical data scarcity.
• Existing T2MedI models may exhibit subgroup biases, neglecting minority groups.
• Investigating and mitigating these biases is crucial for equitable AI in healthcare.

Purpose of the Study:

• To develop and evaluate a T2MedI model with reduced gender bias.
• To propose a novel subgroup distribution alignment method for T2MedI synthesis.
• To ensure generated medical images reflect accurate demographic distributions.

Main Methods:

• Adapted a pre-trained Imagen framework with a fixed Contrastive Language-Image Pre-training (CLIP) text encoder.
• Fine-tuned the decoder on the Radiology Objects in Context (ROCO) dataset.
• Implemented a subgroup distribution alignment method with an alignment loss and CLIP-consistency regularization on the BraTS18 dataset.

Main Results:

• Qualitative and quantitative analyses confirmed gender bias in the initial T2MedI model.
• The proposed alignment method significantly mitigated gender representation inconsistencies in generated brain MR images.
• Generated images better matched the gender distribution of the target BraTS18 dataset.

Conclusions:

• The developed T2MedI approach effectively reduces gender bias in synthetic medical images.
• Subgroup distribution alignment is a viable strategy for creating more equitable AI-generated medical data.
• This work contributes to fairer and more representative medical imaging datasets.