Pneumonia Detection in Veterinary X-Rays Utilizing PaddleSeg-Based Semantic Segmentation and Mamba-Powered Classification.

Pneumonia remains a leading cause of morbidity and mortality in companion animals, with delayed diagnosis significantly impacting treatment outcomes. Current veterinary radiographic analysis primarily relies on manual interpretation, which suffers from substantial inter-rater variability and limited sensitivity for early-stage pathologies. Accordingly, we introduced AI-driven approaches into pet health monitoring, notably outlining a step-by-step diagnostic pipeline to address those above challenges. Key components of the diagnostic pipeline include six parts as follows: data acquisition, data preprocessing, semantic segmentation, classification, risk estimation, and explainability by Gradient-weighted Class Activation Mapping (GradCAM). Utilizing PaddleSeg for semantic segmentation, we accurately localized and extracted the chest area from pet x-ray images, achieving a mean intersection over union (mIoU) of 0.8999. Following this, classification was conducted via the Mamba framework. The model demonstrated robust performance across both frontal (VD/DV) and lateral projections. For frontal (VD/DV) projections, it achieved an accuracy of 90.35%, with a macro-averaged F1-score of 0.9034 (negative: 90.35%; pneumonia: 90.33%) and an area under the ROC curve (AUC) of 0.9438. Performance on lateral projections was equally robust, yielding an accuracy of 89.8%, a macro-averaged F1-score of 0.8965 (negative: 89.64%; pneumonia: 89.66%), and an AUC of 0.9361. Finally, we proposed a novel risk-aware AI module tailored for pet radiographic disease classification, integrating multidimensional uncertainty and interpretability cues into a unified AI Risk Index (ARI), which represents a significant step toward deploying interpretable, safe AI systems in veterinary medicine and sets a precedent for other safety-critical domains.