Detection and classification of eye diseases in cattle using image analysis and deep learning

Abstract

Pinkeye, also known as infectious bovine keratoconjunctivitis (IBK), is a highly contagious ocular disease in cattle characterised by inflammation of the cornea and conjunctiva. It remains one of the most common and economically significant health issues in the livestock industry, yet timely diagnosis remains difficult in Australia’s vast grazing systems, where veterinary access is limited. This thesis presents a deep learning–based diagnostic framework designed to detect and classify pinkeye from mobile phone–captured images of the eye, offering scalable support for on-farm decision-making. A four-part research programme was undertaken. Firstly, You Only Look Oncev5 (YOLOv5) was retrained on 2,000 annotated images from Australia and the USA to localise the eye region in images. Secondly, a 3,800-image dataset was annotated using a clinician developed 17-attribute scorecard and used to train multiple Convolutional Neural Network model with EfficientNetV2B2 consistently outperforming others, achieving high binary classification performance (AUC up to 0.92) and strong ordinal agreement for attributes such as Periocular score (Cohen’s kappa = 0.84). Thirdly, full-stage classification (Normal, Active, Resolving, Resolved) and severity grading were modelled using 3,800 images. While binary treatment classification achieved 94% accuracy, performance dropped for multiclass (69%) and ordinal (κ as low as 0.59) tasks due to certain limiting factors such as sample size, class imbalance, overlapping visual features between disease stages. To enhance transparency, explainable AI tools (Grad-CAM++, LIME, SHAP) were applied to produce heatmaps that reveal the image regions most influential to each prediction, allowing verification that the model’s focus aligns with clinically recognised signs of pinkeye. This thesis demonstrates that deep learning can deliver accurate and interpretable pinkeye diagnostics from field-acquired images under real-field conditions.