<![CDATA[Foot and Mouth Disease (FMD) is a highly contagious livestock disease that causes significant economic losses. Timely detection is essential to prevent rapid transmission. While deep learning has shown promise in image-based disease identification, the impact of integrating lightweight attention mechanisms, such as the Convolutional Block Attention Module (CBAM), into robust multi-scale backbones, such as InceptionV3, for FMD detection on small, imbalanced primary field datasets remains underexplored. This study contributes by providing a systematic evaluation of CBAM integration under varying data-splitting scenarios, highlighting the interaction between attention mechanisms and data distribution. This study evaluates the integration of CBAM into InceptionV3 for the classification of cattle lesion images. It compares its performance with the baseline InceptionV3 model across three train-validation-test splits (70:20:10, 80:10:10, and 70:15:15). The dataset comprises 798 primary images (514 FMD-positive and 284 healthy), indicating a limited size with moderate class imbalance. Images were resized to 299 × 299 pixels and normalized to [-1, 1], with augmentation applied only to the training set. The InceptionV3-CBAM model achieved the best performance under the 70:15:15 split, with 96.69% accuracy, 96.25% precision, 98.72% recall, and 97.48% F1-score. These findings suggest that CBAM can enhance lesion-focused feature representation and detection sensitivity. However, performance gains were inconsistent across splits and appear influenced by both architectural changes and dataset characteristics. The model demonstrates potential for early FMD screening in resource-limited settings, but further validation on larger, more diverse datasets is essential to confirm robustness and generalizability]]>
