RadGazeGen: radiomics and gaze-guided chest X-ray generation using diffusion models

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Text-to-image diffusion models show promise but struggle with clinical accuracy and anatomical fidelity due to limitations of textual descriptions alone.
• Generating clinically accurate medical images requires capturing subtle disease-specific details often missed by text-only prompts.

Purpose of the Study:

• To introduce RadGazeGen, a novel framework for high-fidelity medical image generation.
• To integrate expert eye gaze patterns and radiomic feature maps as controls in text-to-image diffusion models.
• To enhance the anatomical consistency and disease-specific accuracy of generated medical images.

Main Methods:

• RadGazeGen utilizes radiologists' eye gaze trajectories to encode visuo-cognitive attention and spatial localization of disease indicators.
• Radiomic features are employed to capture subvisual phenotypic characteristics like texture, intensity, and shape.
• These multimodal cues (gaze and radiomics) are used as spatial and semantic controls within the diffusion process.

Main Results:

• The framework was evaluated on the REFLACX dataset for image generation quality and diversity.
• Generated images demonstrated high fidelity and diagnostic relevance in downstream tasks, including disease classification on the CheXpert test set.
• Performance was further assessed in long-tailed learning evaluations on the MIMIC-CXR-LT dataset.

Conclusions:

• RadGazeGen successfully bridges the gap between human visual cognition and machine perception in medical image synthesis.
• Jointly conditioning on gaze and radiomic representations improves both the realism and clinical validity of generated medical images.
• The study highlights the importance of anatomically grounded and disease-aware controls for advanced diffusion-based medical image generation.