SeAttnFusionNet for pharyngitis versus no pharyngitis classification from RGB throat images with grad-CAM-based interpretability

Abstract Automated pharyngitis recognition from RGB throat images can support rapid screening and triage, but reliable classification is difficult due to substantial intra-class variability, illumination and viewpoint changes, and strong visual similarity between mild inflammation and healthy appearances. To address these challenges, we propose SeAttnFusionNet, a dual-branch feature extraction network that learns complementary evidence and performs feature-map level fusion before GlobalAveragePooling2D, followed by a compact Dense, Dropout, Dense, Output head. The texture branch refines fine-grained cues using SeparableConv2D, Batch Normalization, and stacked Squeeze-and-Excitation blocks, while the context branch captures structural patterns using Multi-scale convolutions followed by Resize, Concatenate, Normalize, Local Window Attention, a Dynamic Gating Mechanism with Element-wise Multiplication, and Cross Scale Feature Fusion through GAP, Activation, and Sigmoid. The two branch outputs are concatenated and pooled only after fusion, preserving spatially resolved complementary representations prior to global aggregation. We evaluate the model using leakage-safe 5-fold cross-validation with a per-fold 80/20 split on 312 samples consisting of 185 no pharyngitis and 127 pharyngitis cases, with each fold using 250 training samples (148 no pharyngitis, 102 pharyngitis) and a 62-sample test fold (37 no pharyngitis, 25 pharyngitis) at an input size of 224 × 224 pixels. With batch size 32, 50 epochs, Adam optimizer, learning rate 0.0001 scheduled by ReduceLROnPlateau, categorical cross-entropy loss, and regularization via batch normalization and dropout 0.35, SeAttnFusionNet achieves 97.69% train accuracy, 97.26% cross-validation validation accuracy, 97.58% precision, 97.44% recall, and 97.51% F1-score, supported by confusion matrix counts of TN 168, FP 17, FN 3, and TP 124. These results indicate that the proposed fusion of multi-scale attention-driven context modeling with efficient texture refinement provides a practical and reliable framework for automated pharyngitis classification, while Grad-CAM visualizations offer interpretable evidence of discriminative regions without claiming clinical validation.