Enhancing Canine Musculoskeletal Diagnoses: Leveraging Synthetic Image Data for Pre-Training AI-Models on Visual Documentations

The paper addresses the challenge of developing accurate AI-based diagnostic support for canine musculoskeletal conditions when only limited annotated training data exist. In veterinary medicine, large labeled datasets are scarce, making conventional transfer learning from large-scale, dissimilar source tasks (e.g., natural images) less effective. The authors propose mitigating data scarcity by pre-training models on synthetic images that mimic visual body-map documentations of musculoskeletal findings. They generate synthetic datasets (3-class and 36-class variants) and evaluate whether such pre-training improves classification accuracy on a manually curated dataset of five diagnoses, both in a few-shot setting and with more data. The study aims to determine when synthetic pre-training is beneficial for diagnosis accuracy and to support the practical integration of AI into veterinary body-map documentation workflows.

The authors situate their work within several domains: (1) Veterinary informatics faces limited standardized, structured data and often relies on closed practice management systems and free-text notes, hindering AI development (Gysin et al., 2019; Lustgarten et al., 2020; Sheng et al., 2022). (2) Physical and Rehabilitative Medicine (PRM) is growing in veterinary care, with a need for effective documentation and interdisciplinary communication; body maps have been useful in human and animal contexts to represent complex clinical findings (European PRM Bodies Alliance, 2018; Cornwall, 1992; Tarr & Thomas, 2011; Galloway, 2011). (3) Few-shot learning typically transfers knowledge from large-scale, dense-image source tasks, which may not align with sparse, schematic body maps (Hu et al., 2022). (4) Data augmentation spans simple image transformations (Krizhevsky et al., 2012) to deep generative methods like GANs and diffusion models that can synthesize training images (Frid-Adar et al., 2018; Ho et al., 2020; Trabucco et al., 2023), though such models often require large training corpora not available for body maps. Given these constraints, algorithmic generation of synthetic body-map-like images is proposed as a practical alternative.

- Application context and data scarcity: A specialized cross-platform application (built with Flutter) was developed to let veterinarians document findings on standardized body maps and, in the future, integrate with practice management systems. Due to the novelty of body-map-based documentation and the app’s alpha status, few real filled-out maps are currently available, motivating synthetic data generation. - Synthetic data generation (3-class dataset): An algorithm generates synthetic drawings on the body map, forming three classes: (1) line, (2) dashed line, and (3) point cluster. For lines/dashed lines: a random start point s is chosen within the body map and an end point e within a 200-pixel radius, both constrained to lie inside the body map. Curves are drawn as quadratic or cubic Bézier curves with control points deviating randomly up to 30 pixels from the straight line between s and e. For point clusters: the number of points n is sampled uniformly in [3, 20]. A seed point s0 is chosen randomly in the body map, and each subsequent point is sampled within a 20-pixel radius of the previous point, repeated n−1 times. - Synthetic data generation with spatial granularity (36-class dataset): The body map is partitioned into 12 predefined regions. Using the same drawing procedures as above, each of the three primitive classes (line, dashed line, point cluster) is generated but constrained to one of the 12 regions, yielding 36 total classes (3 types × 12 regions). This aims to imbue the model with spatial awareness of the body map. - Evaluation dataset (manual): Five diagnoses with specified visual characteristics were defined: pelvic contusion; atrophy/hypertrophy at the forelimb; atrophy/hypertrophy at the hindlimb; low blood pressure; high blood pressure. For each diagnosis, 50 training and 10 test body maps were manually created (total: 250 train, 50 test). A smaller few-shot subset was also constructed: 5 training and 4 test examples per class (total: 25 train, 20 test). - Model selection and training: Following exploratory comparisons, EfficientNet V2-S (pre-trained on ImageNet) was selected. Two variants were obtained by further pre-training on synthetic data: V2-S-3 (pre-trained on the 3-class synthetic dataset) and V2-S-36 (pre-trained on the 36-class synthetic dataset). All three models (baseline V2-S, V2-S-3, V2-S-36) were then fine-tuned on the real evaluation training sets for the five-way diagnosis classification. Performance was measured as accuracy on the corresponding test sets for both the few-shot subset and the larger dataset.

- Few-shot setting (25 training, 20 test images across five classes): - V2-S (no synthetic pre-training): 80% accuracy - V2-S-3 (pre-trained on 3-class synthetic): 90% accuracy - V2-S-36 (pre-trained on 36-class synthetic): 80% accuracy Result: Synthetic pre-training on the simpler 3-class dataset improved accuracy by approximately 10 percentage points over baseline and 36-class pre-training. - Larger dataset (250 training, 50 test images across five classes): - V2-S: 96% accuracy - V2-S-3: 94% accuracy - V2-S-36: 96% accuracy Result: With more real training data, synthetic pre-training did not improve and in one case slightly reduced accuracy; baseline and 36-class pre-training both achieved 96%. - General observation: More real training images lead to higher accuracy overall. Benefits of synthetic pre-training were evident primarily in the few-shot scenario.

The study shows that synthetic pre-training can improve diagnosis accuracy when only a few labeled examples per class are available, as evidenced by a 10% absolute gain in the few-shot scenario using the 3-class synthetic pre-training. However, when sufficient real training data are available (50 examples per diagnosis), synthetic pre-training provided no advantage and could slightly underperform the baseline. This suggests that synthetic pre-training is most useful for initializing models to the specific visual primitives and spatial context of body maps under data scarcity, but real data dominate as availability increases. The authors note that their synthetic generator captures simple features; more advanced generative approaches (e.g., GANs, diffusion models) may yield more realistic body-map images and potentially improve benefits. They also highlight the potential of incorporating domain knowledge from scientific literature as textual priors to augment learning.

Synthetic pre-training on body-map-like images can mitigate data scarcity in canine musculoskeletal diagnosis tasks, delivering around a 10% accuracy improvement in a few-shot setting. The advantage diminishes with larger real datasets, where baseline or spatially-aware synthetic pre-training performs similarly or better. The work supports the feasibility of body maps as a standardized visual documentation tool and demonstrates that even simple synthetic data can provide meaningful gains under limited data. Future directions include developing more realistic synthetic generators (e.g., GANs/diffusion), integrating textual domain knowledge, expanding diagnoses, and increasing real-world annotated body maps via wider app deployment.

- Limited availability of real filled-out body maps due to the novelty of the documentation method and the application’s alpha status. - Synthetic images capture only simple primitives (lines, dashed lines, point clusters) and may not fully reflect the complexity of clinician-drawn annotations; no GAN/diffusion models were used due to data constraints. - Evaluation covered only five diagnoses with relatively small datasets (few-shot: 25 train/20 test; larger: 250 train/50 test), which may limit generalizability. - Differences between the 3-class and 36-class synthetic pre-training outcomes are not fully explained and warrant further analysis. - Integration with practice management systems was not realized, constraining access to broader, real-world data. - Results suggest benefits mainly in low-data regimes; applicability to diseases with abundant visual documentation is limited.