Domain-specific adaptation for MR image synthesis with text-guided diffusion

Clinical Snapshot

Key Findings

• P Value: Not reported for any outcome

• Effect Size: FID of 13.65 (lower is better; indicates high perceptual realism relative to real images); SSIM of 0.9674 (near-perfect structural similarity); 14% average improvement in Dice Similarity Coefficient for downstream U-Net segmentation

• Primary Outcome: Synthetic MRI image quality and perceptual realism evaluated by quantitative metrics and radiologist blinded assessment; downstream glioma segmentation performance with synthetic data augmentation

• Nnt Or Sensitivity: Radiologist sensitivity for identifying synthetic MRI slices: 25.5% (average across three radiologists); deception rate: 74.5%; 41% of synthetic samples universally misclassified as real by all three radiologists

• Confidence Interval: Not reported for any outcome

Clinical Application

Implementation requires substantial deep learning infrastructure and expertise. The partition-based parallel LDM architecture with Voronoi-grayscale adaptation is computationally complex. Computational requirements, inference time, and hardware specifications are not reported, making direct feasibility assessment impossible. Clinical deployment would require regulatory approval for any downstream diagnostic application. In Australia, this technology is not currently TGA-approved as a medical device software (SaMD) and would require conformity assessment under the TGA's Software as a Medical Device framework before clinical deployment. The RACGP and relevant specialist colleges (RANZCR) would require robust clinical validation studies before endorsing synthetic data augmentation in diagnostic imaging AI pipelines. The method may have relevance to Australian rare disease research initiatives and the Australian Genomics Health Alliance's neurological tumour programs, where small dataset sizes are a recognised constraint. PBS implications are indirect — improved segmentation AI could support radiotherapy planning workflows relevant to Medicare-funded services. Any use of synthetic patient-derived imaging data would need to comply with the Australian Privacy Act 1988 and relevant state health records legislation. Research and development contexts involving small-scale glioma MRI datasets where data augmentation is required for training deep learning segmentation or classification models; potential extension to other rare neurological tumours with limited imaging data

Abstract

Objective.Deep learning in medical imaging is severely constrained by data scarcity. Data synthesis offers a promising solution, but existing generative models have difficulty in restoring pathological texture features when trained on small-scale datasets. To address this, we propose a domain-specific, partition-based parallel text-guided latent diffusion model (LDM) for medical image synthesis.Approach.Each LDM operates on a defined image domain and is fine-tuned to reproduce specific texture characteristics. Diseased regions are identified from segmentation masks, while healthy regions are further subdivided using Voronoi-grayscale adaptation, enabling localized texture preservation. The fine-tuned LDMs independently synthesize corresponding image partitions, which are subsequently merged and denoised to form complete synthetic images with paired segmentation masks.Main results.We evaluated the approach on glioma MRI data, achieving a Fréchet Inception Distance of 13.65, demonstrating high perceptual realism. Texture fidelity was further supported by SSIM of 0.9674 and radiomic feature distribution analysis, both confirming close alignment between real and synthetic images. In a blinded visual Turing test, three radiologists achieved an average sensitivity of only 25.5% when identifying synthetic MRI slices, resulting in a 74.5% deception rate, and 41% of the synthetic samples were universally misclassified as real by all experts. In downstream experiments, U-Net trained on the synthetic-augmented dataset improved DSC by 14% on average.Significance.These results demonstrate that the proposed domain-specific adaptation framework can generate perceptually plausible, structure-preserving synthetic MRI slices in data-constrained environments, while improving downstream segmentation performance. The method therefore shows potential as an augmentation-oriented tool for AI model development, clinical teaching, assisted diagnosis, and rare-disease research.

References

1. 1.Wen, Y., Healy, J. J., Song, Y., Fu, X., Song, S., & Curran, K. M. (2026). Domain-specific adaptation for MR image synthesis with text-guided diffusion. Physics in Medicine and Biology. https://doi.org/10.1088/1361-6560/ae7797