This paper investigates the use of synthetic image data for pre-training AI models to improve the accuracy of canine musculoskeletal diagnoses. Due to the novelty of a new visual documentation method (body maps), training data is scarce. The researchers generated synthetic datasets (one with 3 classes, another with 36) mimicking realistic visual documentation. An AI model was pre-trained on these datasets and then evaluated on a manually created evaluation dataset (250 examples). Results showed a 10% accuracy improvement on a smaller subset (25 examples), but this improvement wasn't observed on the larger dataset. The study suggests that synthetic data is beneficial primarily when training data is very limited.

Martin Thißen, Thi Ngoc Diep Tran, Ben Joel Schönbein, Ute Trapp, Barbara Esteve Ratsch, Beate Egner, Romana Piat, Elke Hergenröther