PharynFuseFormer supports image based pharyngitis screening from throat photographs using transformer CNN feature fusion

Pharyngitis is a common cause of throat-related clinical presentation, yet image-based screening from throat photographs remains insufficiently standardized and often lacks rigorous, leakage-aware evaluation. Existing approaches are commonly limited by small datasets, single-representation models, and incomplete validation protocols, which can weaken confidence in reported generalization. To address this gap, this study proposes PharynFuseFormer, a hybrid transformer–CNN feature-fusion framework that combines multi-scale convolutional processing, attention-based contextual refinement, and channel-wise gating to jointly model local texture cues and broader spatial context for binary pharyngitis screening. Experiments were conducted on a curated dataset of 362 RGB throat images, comprising 147 Pharyngitis and 215 Non-Pharyngitis samples, with a development split of 329 images and a held-out test split of 33 images. Images were resized to 224 × 224 × 3, and augmentation was applied only within the development split to reduce leakage risk. Evaluation used stratified 5-fold cross-validation with statistical comparison against representative baselines. PharynFuseFormer achieved 99.10% training accuracy and 98.73% cross-validation accuracy, with an improvement of up to + 4.73% points (95% CI + 4.18 to + 5.28; Holm p = 1.06e − 05; Cohen’s d = 14.01). Qualitative analyses using Grad-CAM, final-layer feature maps, and averaged activation heatmaps were included to contextualize model behavior. Under the current experimental setting, these findings support transformer–CNN feature fusion as a promising approach for image-based pharyngitis screening, while further validation under diverse acquisition conditions remains necessary.