SRA-Seg: Synthetic to Real Alignment for Semi-Supervised Medical Image Segmentation

Abstract

Synthetic data, an appealing alternative to extensive expert-annotated data for medical image segmentation, consistently fails to improve segmentation performance despite its visual realism. The reason being that synthetic and real medical images exist in different semantic feature spaces, creating a domain gap that current semi-supervised learning methods cannot bridge. We propose SRA-Seg, a framework explicitly designed to align synthetic and real feature distributions for medical image segmentation. SRA-Seg introduces a similarity-alignment (SA) loss using frozen DINOv2 embeddings to pull synthetic representations toward their nearest real counterparts in semantic space. We employ soft edge blending to create smooth anatomical transitions and continuous labels, eliminating the hard boundaries from traditional copy-paste augmentation. The framework generates pseudo-labels for synthetic images via an EMA teacher model and applies soft-segmentation losses that respect uncertainty in mixed regions. Our experiments demonstrate strong results: using only 10% labeled real data and 90% synthetic unlabeled data, SRA-Seg achieves 89.34% Dice on ACDC and 84.42% on FIVES, significantly outperforming existing semi-supervised methods and matching the performance of methods using real unlabeled data.

Method

Figure: Synthetic data generated by StyleGAN2-ADA for ACDC (top) and FIVES (bottom) datasets. Left to right: first 3 are original images, next 3 generated using 5% real data, last 3 using 10% real data.

Domain Gap Visualization

(a) UNet + Real Unlabeled

(b) UNet + Synthetic

(c) BCP + Synthetic

(d) SRA-Seg (Ours)

Figure: KDE plots showing domain mismatch between labeled (green) and unlabeled (blue) data. Our method (d) effectively aligns the distributions.

Figure: Soft Edge Blending component of SRA-Seg. The labeled and unlabeled images as well as the corresponding segmentation masks are mixed through cropping and soft mask blending to reduce sharp edges.

Results

ACDC Dataset (Cardiac MRI)

Bold is best and underline is second best, ranked within the synthetic-unlabeled block of each split. The real-unlabeled rows are reference upper bounds and are not ranked.

FIVES Dataset (Fundus Images)

Bold is best and underline is second best, ranked within the synthetic-unlabeled block of each split. The real-unlabeled rows are reference upper bounds and are not ranked.

Qualitative Results

ACDC Dataset (10% labeled, 90% synthetic). Incorrectly segmented pixels are highlighted in red. Click any image to enlarge.

Image	UNet	BCP	CrossMatch	ABD	DiffRect	CGS	SRA-Seg	GT

FIVES Dataset (10% labeled, 90% synthetic). Incorrectly segmented pixels are highlighted in red. Click any image to enlarge.

Image	UNet	BCP	CrossMatch	ABD	DiffRect	CGS	SRA-Seg	GT

Figure: Comparison of real-image usage (bar height) and resulting Dice scores (red markers) for BCP versus SRA-Seg on ACDC and FIVES datasets. SRA-Seg achieves higher Dice with only 10% real data.

Figure: FID scores comparing synthetic and real images across 5% and 10% data splits. Lower FID indicates better synthetic data quality. FIVES achieves consistently lower FID scores.

Ablation Study

All configurations use 10% real labeled and 90% synthetic unlabeled data. The first row is the BCP baseline.

Bold is best per metric. Soft-mix and soft-loss together give the largest Dice gain; SA-loss adds the final improvement and is what closes the distribution gap shown in the KDE plots.

BibTeX

@InProceedings{aranya2027sraseg,
  author    = {Aranya, OFM Riaz Rahman and Desai, Kevin},
  title     = {SRA-Seg: Synthetic to Real Alignment for Semi-Supervised Medical Image Segmentation},
  booktitle = {Pattern Recognition (ICPR)},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  pages     = {698--715},
  year      = {2027},
  doi       = {10.1007/978-3-032-31335-5_47},
  isbn      = {978-3-032-31335-5}
}

Acknowledgments

This work builds upon BCP and utilizes StyleGAN2-ADA for synthetic data generation and DINOv2 for feature extraction.