PSIV-10 Towards automated anemia detection: AI models for accurate FAMACHA classification in outdoor environments.

Abstract Accurate classification of FAMACHA images is essential for effective anemia detection in small ruminants. However, environmental conditions such as lighting variations (daylight vs. shade) can influence image quality and diagnostic accuracy. This study aimed to evaluate the performance of Support Vector Machine (SVM) and Backpropagation Neural Network (BPNN) models in classifying FAMACHA images taken in daylight (No Shade = 1) and shaded areas (Shade = 0). A dataset of 1000 images (500 per group) was analyzed using machine learning models, with five different data augmentation techniques applied to enhance robustness and using 10 10-fold nested cross-validation techniques. Model performance was evaluated based on accuracy, precision, recall, F1-score, and Cohen’s Kappa statistic to measure classification agreement. The SVM model achieved an overall accuracy of 98%, with a precision and recall of 0.98 for both shaded and non-shaded images. The BPNN model obtained an accuracy of 97%, with precision ranging from 0.94 to 1.00 and recall values from 0.94 to 1.00. Cohen’s Kappa values were 0.959 for SVM and 0.939 for BPNN, indicating strong agreement between predicted and actual classifications. Both models demonstrated high reliability, though SVM performed slightly better in handling lighting variations. The SVM model excelled in balancing classification across both conditions, while BPNN exhibited minor inconsistencies in differentiating shaded images due to slight recall fluctuations. The findings indicate that SVM is the superior model for FAMACHA image classification, particularly in outdoor settings where lighting conditions vary. The high Cohen’s Kappa scores suggest strong inter-model agreement, reinforcing the effectiveness of machine learning in automating FAMACHA classification. Incorporating data augmentation enhanced model generalizability, reducing the risk of misclassification due to lighting inconsistencies. Further research will explore the integration of ensemble learning techniques, combining the strengths of SVM and BPNN to enhance classification robustness. Expanding the dataset to include diverse environmental conditions and using deep learning approaches such as convolutional neural networks (CNNs) may improve performance further. The development of mobile-based AI applications for real-time, automated FAMACHA assessment would significantly benefit precision livestock farming. By leveraging AI-powered image classification, livestock health monitoring can become more efficient, scalable, and accessible, reducing reliance on manual assessments and improving parasite management strategies in small ruminant farming.