Robust and explainable poultry disease classification via MaxViT with attention-guided visualization

Poultry farming plays a crucial role in global food security, yet it faces significant challenges from infectious diseases such as Coccidiosis, Salmonella, and Newcastle disease, which can lead to substantial economic losses. Existing deep learning models often encounter issues like class imbalance, poor generalization across different datasets, and a lack of interpretability. To address these limitations, we propose a novel hybrid deep learning framework based on the MaxViT architecture. This framework combines MBConv blocks with both block-wise and grid-based self-attention mechanisms, allowing it to effectively capture local and global features in complex fecal images. In our study, we utilized two publicly available poultry fecal image datasets, consisting of 8,067 and 6,812 images, respectively. Each dataset includes four classes: Coccidiosis, Salmonella, Newcastle disease, and Healthy. To tackle the severe class imbalance, we applied data augmentation techniques. We evaluated the models using various metrics, including accuracy, F1-score, specificity, PR AUC, and MCC, under 10-fold cross-validation. Our proposed MaxViT model achieved impressive accuracy scores of 99.54% and 98.96% on the two datasets, outperforming ViT-B/16, ViT-L/32, DeiT-S, and T2T-ViT-14 across all metrics. Additionally, we integrated Grad-CAM to provide visual explanations of the model's decisions, thereby enhancing transparency and applicability in veterinary settings. This study introduces a deployable, interpretable, and highly accurate framework for intelligent poultry disease diagnosis, effectively addressing critical limitations found in previous research.