Source-linked AI summary
Synthetic-to-Real Transfer in Cerebral Microbleed Generation and Segmentation
To-Liang Hsu, Ting-Yu Lai, Ching-Ting Lin, Chun-Hao Huang, Wei-Chun Wang
TL;DR
Low CMB prevalence and costly expert annotation constrain automated detection training, and architecture-dependent synthetic-to-real transfer has been little explored. This study generates synthetic CMBs and compares training paradigms across two detection frameworks. Synthetic training transfers substantially but trails real training, and transfer varies by framework.
Problem
CMBs are rare, difficult to distinguish from mimics, and costly to annotate, while architecture-dependent synthetic-to-real transfer remains largely unexplored.
Source: §1 · p. 2Method
The authors generate synthetic CMBs on CMB-negative scans and compare real, synthetic, and mixed training paradigms across MONAI 3D U-Net and nnU-Net.
Source: §1 · p. 2Results
Synthetic training transferred substantially to real CMBs but real-data training consistently outperformed it; MONAI S75 retained 87.7% of the lesion sensitivity achieved by MONAI R78, while transfer differed substantially between frameworks.
Source: §4.1 · p. 14, §5 · p. 22Takeaways & Limitations
Synthetic lesions provide useful supervision on real images, but did not fully replace real positive cases in these experiments, and transfer depended on the detection framework.
Source: §5 · p. 22Only two downstream frameworks were evaluated, and the experiments do not identify which model or training characteristics account for their transfer differences.
Source: §5 · p. 23
Abstract
from arXiv · showhide
The development of automated cerebral microbleed (CMB) detection models is hindered by the low prevalence of CMBs and the high cost of expert annotation. To address this limitation, we developed a synthetic CMB generation pipeline and investigated the effectiveness of synthetic lesions for training deep learning detectors. Models trained solely on synthetic data achieved substantial detection performance, reaching approximately 87% of the lesion sensitivity of their real-trained counterparts. We further investigated the complementary roles of synthetic and real data under different training paradigms, revealing that a significant performance gap remains. Moreover, we found that successful synthetic-to-real transfer is strongly dependent on the downstream detection architecture, providing new insight into both the potential and limitations of synthetic data for CMB detection.
1 Introduction
CMB detection is difficult because lesions are rare, small, and easily confused with vessels or artifacts, while expert annotation is costly. Prior synthetic-data studies left open whether transfer depends on detector architecture; this work compares training paradigms across two frameworks.
1 Introduction
- CMBs are sparse and can resemble vessels or imaging artifacts, while expert annotation is expensive and time-consuming.CMBs are associated with small vessel disease and are imaging biomarkers linked to hemorrhage, stroke, cognitive impairment, and dementia.Source: §1 · pp. 1–2
- Earlier synthetic CMB studies used one downstream detection framework, leaving architecture-dependent synthetic-to-real transfer largely unexplored.One prior approach yielded approximately nine false positives per scan in whole-brain detection; another mainly evaluated synthetic lesions alongside real training data.Source: §1 · p. 2
- The study compares four real/synthetic training paradigms across MONAI 3D U-Net and nnU-Net to examine synthetic-to-real transfer and detector design.Synthetic targets are generated on CMB-negative scans, separating synthetic training data from real positive lesions.Source: §1 · p. 2
2 Data
The study uses heterogeneous clinical MRI examinations, manually labeled lesion masks, and a partition separating training from final hold-out evaluation.
2.1 Data Acquisition
- 205 MRI examinations comprised 104 CMB-positive and 101 CMB-negative cases, acquired on multiple scanner models at 1.5T and 3T.All examinations used SWAN sequences in routine clinical practice.Source: §2.1 · p. 3
2.2 Microbleed Annotation
- Annotations included lesions 2.0–10.0 mm in diameter, manually labeled as voxel-wise masks on SWAN images by an experienced neurologist.CMB-negative examinations were verified to contain no visible CMBs and served as negative controls.Source: §2.2 · pp. 3–4
2.3 Dataset Partitioning
- The hold-out test set was reserved for final evaluation, while 78 CMB-positive exams trained real-data models and 75 CMB-negative exams served only as synthetic-lesion hosts.Positive subjects were stratified by lesion burden when assigned to preserve the distribution across sets.Source: §2.3 · p. 4
3 Methods
The methods standardize SWAN images, generate parameterized synthetic CMBs in CMB-negative scans, and evaluate them across segmentation frameworks, training paradigms, and evaluation levels. Synthetic lesions approximate real-lesion distributions, with some characteristics remaining moderately different.
3.1 Overview
- Table 2 summarizes the distribution of CMB-positive cases by lesion burden.Source: Table 2 · p. 4
3.2 Preprocessing
- All datasets underwent standardized preprocessing, including DICOM-to-NIfTI conversion, brain extraction with a dedicated nnU-Net v2 model, and skull-stripped image normalization to 0–255.The brain-extraction model was trained on 20 delineated SWAN volumes from the CMB-positive training cohort.Source: §3.2 · p. 5
3.3 Synthetic Microbleed Generation Pipeline
- Synthetic CMBs were modeled as ellipsoids with independently sampled size and elongation, inserted uniformly within an eroded brain mask, and blended as hypointense lesions with gradual boundaries.The generator used real-positive training-cohort measurements; five lesions were inserted per healthy MRI, approximately matching the training cohort’s median burden.Source: Synthetic Lesion Generation · pp. 6–8
- Synthetic-lesion comparisons covered burden, volume, intensity, contrast, and elongation, with ECDFs and Kolmogorov–Smirnov statistics used to characterize distributional differences.The provided figure captions describe the visualized characteristics but do not state additional comparison outcomes.Source: Comparison of Real and Synthetic… · p. 8, Figure 6 · p. 9, Figure 7 · p. 10
- Synthetic lesions reproduced the central tendency and distributions of volume and median intensity relatively closely, while elongation and background-to-lesion contrast remained moderately different.The comparison characterizes similarity, not exact distributional equivalence.Source: Comparison of Real and Synthetic… · p. 9, Table 4 · p. 11
3.4 Detection Network
- The primary detector was a MONAI 3D U-Net, with nnU-Net v2 included as an unmodified self-configuring framework to test architecture-dependent transfer.The MONAI network produced voxel-wise probability maps from single-channel SWAN images.Source: §3.4 · p. 10, MONAI 3D U-Net · p. 12, nnU-Net · p. 13
3.5 Training Paradigms
- Training paradigms compared nested real and synthetic subsets, synthetic training with real-data validation, and approximately 1:1 mixed real-synthetic datasets.Subset sizes were 20, 40, and 78 real examinations and 20, 40, and 75 synthetic examinations.Source: §3.5 · p. 13
3.6 Evaluation Metrics
- Performance was assessed on an independent hold-out set at voxel, lesion, and case levels, including Dice, lesion sensitivity, false positives per case, and case-level F1.Lesion detection required any voxel overlap; case positivity required at least one predicted lesion after component filtering.Source: §3.6 · pp. 13–14
4 Results
Across four training paradigms and two detection frameworks, the experiments show that synthetic CMBs can transfer to real lesions, but results vary by framework and real-data training remains stronger across the tested training sizes. Adding real data through validation or synthetic data through mixed training did not yield a clear overall improvement across metrics.
4.1 Pure Synthetic-To-Real Transfer
- MONAI S75 retained 83.7% of MONAI R78’s mean Dice and 87.7% of its lesion sensitivity, despite no real CMB annotations during training or validation.The transfer came with a higher false-positive rate: 1.423 versus 1.134 FP/case.Source: §4.1 · pp. 14–15, Table 5 · p. 15
4.2 Effect of Training Set Size
- Mean Dice was 0.660 and lesion sensitivity was 0.807 for MONAI R20, exceeding MONAI S75 at Dice 0.576 and lesion sensitivity 0.749, and all other synthetic-only configurations.Real-data training produced higher Dice and lesion sensitivity than synthetic-only training across the evaluated training set sizes.Source: §4.2 · p. 15, Table 6 · p. 18
4.3 Incorporating Real and Synthetic Data
- MONAI S75+ReVal showed no clear overall improvement over S75, while mixing synthetic data with real training increased lesion sensitivity but lowered Dice and raised FP/case.For R78+S75, lesion sensitivity rose from 0.854 to 0.877 while FP/case rose from 1.135 to 1.788.Source: §4.3 · p. 17, Table 6 · p. 18
4.4 Synthetic-to-Real Transfer Across Detection Frameworks
- MONAI S75 reached lesion sensitivity 0.749 versus 0.854 for MONAI R78, while nnUNet S75 reached 0.269 versus 0.860 for nnUNet R78, showing markedly different synthetic-to-real transfer across frameworks.The comparisons use each framework’s real-only model as its reference.Source: §4.4 · p. 17, Table 6 · p. 18
5 Discussion
Synthetic lesions inserted into real CMB-negative scans provided useful supervision, but real-data training remained stronger in the tested settings. Transfer varied by detection framework, and the experiments do not identify the mechanism behind that difference.
5 Discussion
- Synthetic CMBs provided useful supervision on real images, but real-data training consistently outperformed synthetic-data training across the evaluated data scales.The authors caution that this gap is not an inherent limit of synthetic training because only one generation setting was evaluated.Source: §5 · p. 22
- Transfer differed between MONAI 3D U-Net and nnUNet, but the experiments do not identify its mechanism because the frameworks differ in several characteristics.Model capacity, preprocessing, augmentation, and training procedure may each contribute.Source: §5 · p. 22
- The generator models selected CMB characteristics and leaves distributional differences, including background-to-lesion contrast; other lesion shapes, boundaries, and imaging artifacts remain possible future targets.Only one synthetic generation setting was used across the downstream experiments.Source: §5 · p. 22
- The authors caution against assuming this approach generalizes to more complex pathologies, since CMBs are small, localized lesions with relatively simple visual characteristics and limited effects on surrounding anatomy.Those characteristics make direct insertion into CMB-negative scans feasible in this study.Source: §5 · p. 22