Sheep pulmonary image recognition (SPIRO): An explainable deep learning framework for macroscopic assessment of ovine pulmonary lesions

Accurate detection and classification of pulmonary diseases in sheep are critical for ensuring animal welfare, food safety, and effective disease surveillance in abattoir settings. In this study, we propose a two-stage, explainable, histopathology-informed deep learning framework for the automated classification of ovine lung lesions. The dataset comprises 1,052 sheep lung images, annotated through a dual-level diagnostic process involving macroscopic inspection and histopathological confirmation, and classified as healthy, inflammatory, or neoplastic. Four convolutional neural network architectures-VGG16, VGG19, ResNet50, and EfficientNetB0-were benchmarked under identical conditions. EfficientNetB0 achieved the highest observed standalone performance and was integrated into a two-stage pipeline, in which YOLOv11 first localized lung regions and EfficientNetB0 subsequently classified the cropped images. Model interpretability was evaluated using Gradient-weighted Class Activation Mapping (Grad-CAM). The YOLOv11-EfficientNetB0 pipeline achieved an overall test accuracy of 90% and a weighted F1-score of 0.88. Neoplastic lesions were classified with the highest performance, reaching precision, recall, and F1-score values of 0.98, 0.99, and 0.99, respectively, while healthy and inflammatory lungs achieved F1-scores of 0.83 and 0.82, respectively. Grad-CAM visualizations indicated that predictions were mainly driven by diagnostically relevant pulmonary regions. These findings demonstrate the feasibility of explainable artificial intelligence for objective, scalable, and biologically grounded post-mortem assessment of sheep lungs, with potential applications in abattoir surveillance, veterinary pathology, and flock health monitoring.