Deep learning for pediatric synovial recess distension detection in hemophilia: synthetic image augmentation with styleGAN2-ADA.
Authors
Affiliations (8)
Affiliations (8)
- Department of Medical Imaging, University of Toronto, 263 McCaul St, 4 Floor, Toronto, ON, M5T 1W7, Canada.
- Children's Hospital of Central Switzerland, Kinderspital Zentralschweiz Und Kantonsspital Luzern, 6000, Spitalstrasse, Luzern 16, Switzerland.
- German Center for Pediatric Rheumatology, Gehfeldstrasse 24, 82467, Garmisch-Partenkirchen, Germany.
- Rheumatology Center Rhineland Palatinate, Hospital for Pediatric Rheumatology, Kaiser-Wilhelm-Str. 9-11, 55543, Bad Kreuznach, Germany.
- Department of Hematology and Oncology, Kansai Medical University Hospital, 2 Chome-3-1 Shinmachi, Hirakata, Osaka, 573-1191, Japan.
- Department of Medical Imaging, University of Toronto, 263 McCaul St, 4 Floor, Toronto, ON, M5T 1W7, Canada. [email protected].
- Institute of Medical Science, 1 King's College Circle, University of Toronto, Toronto, ON,, M5S 3K3, Canada. [email protected].
- Department of Statistical Sciences, University of Toronto, 700 University Ave, 9 Floor, Toronto, ON, M5G 1X6, Canada. [email protected].
Abstract
Pediatric musculoskeletal ultrasound (MSKUS) datasets are scarce, especially for rare, sex-linked conditions such as hemophilia. Models trained on adult data have limited generalizability due to anatomical differences. We aimed to improve pediatric synovial recess distension (SRD) classification by augmenting real data with synthetic images. We developed a tailored augmentation framework using conditional StyleGAN2-ADA to generate age- and diagnosis-specific synthetic ultrasound images (0-8, 9-13, 14-18 years; SRD-positive/negative). Our two-stage quality control pipeline (distance-based filtering and Support Vector Machine (SVM) confidence weighting) using task-specific EfficientNet-B4 embeddings ensured anatomical plausibility. The dataset-2,499 real pediatric and adult knee ultrasound images and 21,550 quality-controlled synthetic images-was used to fine-tune an EfficientNet-B4 classifier, evaluated on an independent pediatric test set of 278 images across four ablation configurations. A separate 3-class age classifier validated anatomical feature preservation. Statistical comparisons used McNemar's test, per-fold sign tests, and bootstrap confidence intervals. Our proposed model improved accuracy over the adult baseline by + 17.3 pp for ages 0-8 (85.3% vs. 68.0%, p < 0.001) and + 6.1 pp for ages 14-18 (89.9% vs. 83.8%, p = 0.033), with consistent gains across all five cross-validation folds (sign test p = 0.031). An independent age classifier confirmed that quality-controlled synthetic images preserved age-specific anatomical features (macro accuracy 0.844 vs. 0.644 real-only). Conditional StyleGAN2-ADA with two-stage quality control improved pediatric SRD classification and preserved age-specific anatomical relevance, supporting accurate, age-aware AI tools for rare pediatric conditions.