Cross-Species Disease Detection Model Using Domain Adaptation

Veterinary radiology faces persistent hurdles for deep learning: limited labeled data within each species and substantial domain shift driven by anatomical, acquisition, and contrast differences. We investigate a domain adaptation framework that transfers a pneumonia detector trained on canine chest radiographs to feline radiographs, enabling accurate, dataefficient cross-species diagnosis without requiring large labeled target datasets. The approach integrates adversarial distribution alignment with optional semi-supervised fine-tuning, and supports deployment practices such as probability calibration and visual explanations.