CNN and LSTM architectures achieve up to 99% accuracy in veterinary diagnostic decision supportNew AI models show high accuracy in diagnosing pet illnesses

Early disease detection prior to the onset of advanced clinical signs remains a critical imperative in veterinary medicine. This systematic review provides a methodological synthesis of Convolutional Neural Networks and Long Short-Term Memory networks applied to diagnostic and clinical decision support in canine and feline patients, evaluating both algorithmic performance and clinical feasibility. Adhering to the PICO-CS framework and PRISMA 2020 guidelines, we queried Scopus and Web of Science for peer-reviewed studies published between 2021 and 2026, retrieving a final corpus of 57 primary studies. Eleven clinical domains were identified, with evidence heavily concentrated in Oncology (28.1%), Cardiology (15.8%), and Pulmonology (12.3%). CNN architectures predominated in spatial imaging modalities—including radiographs, clinical photography, ultrasonography, and microscopic imagery—whereas LSTM networks were deployed for temporal signals and sensor-derived telemetry. Binary classification tasks achieved the highest performance metrics (accuracy reaching up to 99%), whereas multi-class formulations exhibited broader variability due to dataset imbalance. Cross-study benchmarking revealed substantial heterogeneity in datasets, evaluation metrics, and validation protocols, while reliance on high-performance computational infrastructure poses a barrier to point-of-care deployment in resource-constrained clinics. We conclude that while deep learning architectures hold significant promise as clinical decision-support tools, transitioning to routine veterinary practice requires prospective external validation, representative open-access datasets, and lightweight models optimized for conventional clinical hardware.