Supporting transformer-based cardiac MRI segmentation with text-to-image controllable diffusion pipelines

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cardiovascular Diseases

Background:

• Cardiac MRI is vital for diagnosing cardiovascular diseases.
• Acquiring large, annotated cardiac MRI datasets is challenging due to ethical, economic, and logistical constraints.

Purpose of the Study:

• To propose a generative framework using diffusion models for controllable synthesis of anatomically consistent cardiac MRI scans with segmentation labels.
• To enable the creation of fully annotated synthetic datasets for training AI segmentation models without manual labeling.

Main Methods:

• Utilized diffusion models combined with Low-Rank Adaptation (LoRA) for pathology-aware label map generation from text prompts.
• Employed ControlNet for semantic and spatial conditioning to guide image synthesis.
• Evaluated image realism using FID, KID, and FRD metrics.
• Assessed downstream segmentation performance using a SegFormer model trained on real, synthetic, and combined datasets.

Main Results:

• The proposed framework significantly improves segmentation accuracy, especially in data-limited scenarios.
• Cross-dataset experiments on the M&Ms cohort demonstrated improved generalization and robustness to domain shifts.
• Generated synthetic data aligned with cardiac pathologies and phases, supporting supervised training.

Conclusions:

• Text-guided diffusion pipelines can generate high-quality, semantically consistent medical imaging data.
• This approach offers a viable solution for robust and generalizable AI training in medical imaging, overcoming data acquisition challenges.