Shrimp aquaculture is a critical source of global food production but faces grave challenges posed by infectious diseases such as Black Gill (BG) disease and White Spot Syndrome Virus (WSSV). In this work, we utilized a deep learning approach to multi-class classification of the health condition of shrimp samples into classes Healthy, BG, WSSV, and co-infected (BG_WSSV). We evaluated three top convolutional neural network models, InceptionV3, Xception, and a Dual-Xception fusion model, on 5-fold cross-validation, each trained for more than 10 epochs with default parameters and equal computational power. Performance metrics like accuracy, precision, F1 score, TPR, TNR, FPR, and FNR were thoroughly analyzed. The Dual-Xception outperformed with more than 87% accuracy on more than one-fold with special preference in detecting co-infection. Grad-CAM visualizations were used to model focus interpretation and to cross-check feature learning, where feature learning was verified as the model can identify biologically relevant regions such as gills and abdominal areas. The results depict an encouraging deep learning model for interpretable and accurate disease screening of shrimp aquaculture.