The application of AI in veterinary medicine is hampered by the limited availability of large, annotated datasets. This is particularly true for specialized areas like canine musculoskeletal diagnosis. While few-shot learning (FSL) offers a potential solution, its effectiveness is reduced when the target task differs significantly from existing large-scale datasets. To address this, the researchers explored data augmentation using synthetic examples mimicking real-world visual documentation of canine musculoskeletal conditions. The study aimed to determine if pre-training an AI model with synthetic data improves the accuracy of diagnoses. This approach addresses the challenge of data scarcity in veterinary AI by generating synthetic images that reflect the characteristics of real-world visual documentation of canine musculoskeletal diseases. The goal was to create a system that could assist veterinarians in making more accurate diagnoses.

The paper cites existing literature highlighting the challenges of limited data in veterinary AI (Appleby and Basran 2022) and the potential of few-shot learning (Hu et al. 2022). It also mentions previous work on data augmentation using synthetic data (Trabucco et al. 2023, Burg et al. 2023) and the use of GANs and diffusion models for generating synthetic medical images (Frid-Adar et al. 2018, Trabucco et al. 2023). The authors note that most veterinary practices use practice management systems (PMS) which are often closed systems, making data access challenging (Gysin et al. 2019). The lack of standardization in veterinary record-keeping is also mentioned, with most information stored as free text (Lustgarten et al. 2020). Some research has explored extracting information from free text for classification purposes (Sheng et al. 2022). The authors mention existing data augmentation techniques such as horizontal flipping, color space augmentation, and random cropping (Krizhevsky et al. 2012), but they are not suitable for standardized, digitally created body maps.

The researchers developed a method to generate synthetic images mimicking visual documentation of canine musculoskeletal diagnoses using body maps. Initially, a basic dataset with three classes (line, dashed line, point cluster) was created. These represent different severities of clinical findings. The algorithm randomly places these elements within the body map. This was extended to a more sophisticated dataset with 36 classes by dividing the body map into 12 regions and generating the three classes within each region. This aimed to introduce spatial awareness. An evaluation dataset of 250 manually created visual documentations for five different diagnoses (50 examples per diagnosis) was also created, with a subset of 25 examples used for few-shot learning evaluation. The EfficientNet V2-S model (pre-trained on ImageNet) was chosen as the classification model after exploratory analysis. Two versions of the model were created (V2-S-3 and V2-S-36) by fine-tuning it on the generated synthetic datasets. These were then further fine-tuned on the evaluation dataset's training data, along with the original V2-S model. The models were evaluated on their accuracy on both the full evaluation dataset and the smaller subset.

The results demonstrated a 10% improvement in accuracy (from 80% to 90%) for the V2-S-3 model (pre-trained on the 3-class synthetic dataset) on the smaller evaluation dataset (25 examples). However, this improvement was not observed on the larger evaluation dataset (250 examples), where the V2-S and V2-S-36 models achieved 96% accuracy, while V2-S-3 achieved only 94%. This suggests that pre-training with synthetic data is beneficial mainly in low-data scenarios. The use of synthetic data did not universally improve the accuracy across all datasets and model configurations. The increased accuracy in the smaller dataset highlights the utility of synthetic data augmentation when actual data is scarce.

The findings indicate that synthetic data can improve the accuracy of canine musculoskeletal diagnoses when training data is limited, particularly in few-shot learning scenarios. The lack of improvement on the larger dataset suggests that the benefit of synthetic data diminishes as the amount of real-world training data increases. The study highlights the potential of using synthetic data to address data scarcity issues in veterinary AI. While a simple algorithmic approach was used for synthetic data generation, more sophisticated methods such as GANs or diffusion models could potentially generate more realistic images and lead to further improvements. The integration of textual data (scientific literature) as prior knowledge is suggested as a future research direction.

This work demonstrates the potential of synthetic data augmentation to improve the accuracy of AI-based diagnostic support systems for canine musculoskeletal diseases, especially when real-world training data is limited. While the benefits diminished with larger datasets, the results showcase the utility of synthetic data in low-data scenarios. Future work should explore more advanced generative models and the incorporation of textual data as prior knowledge.

The study's limitations include the relatively simple algorithmic approach used for generating synthetic data. More complex models might generate more realistic images. The specific characteristics and number of classes in the synthetic datasets might influence the results. Further research is needed to fully understand the optimal balance between real and synthetic data for training AI models in this domain.