Automated Pharyngitis Classification Using a Dual-Branch Hybrid Deep Learning Model with Explainable Artificial Intelligence and Retrieval-Augmented Generation-Based Clinical Recommendation

Abstract

Pharyngitis is a frequent inflammatory condition of the upper-airway system for which prompt preliminary evaluation can aid in the proper clinical evaluation and minimize treatment outside of that. However, the level of inflammation, the quality of illumination, the quality of the image, and the level of the observer’s experience can affect the interpretation of the image. A research-stage framework for binary pharyngitis screening with 312 pharyngitis images classified as no pharyngitis and pharyngitis is presented. The proposed dual-branch hybrid deep learning model consists of a separable convolution-based local feature extraction branch and an attention-based global contextual learning and cross-scale feature fusion branch. It brings together the combined power of hybrid CNN/attention based classification, explainable AI, and retrieval-augmented generation, along with a Qwen large language model, in a single screening workflow. The model performance was done by 5-fold cross-validation with leakage-safe and the result is 97.25% accuracy, 97.58% precision, 97.43% recall, and 97.50% F1 score. To explain the predictions, image regions were emphasized by explainability visualization, and the RAG/Qwen module produced structured recommendation reports using the predicted class and retrieved medical context. The framework can help with initial image screening, and clinician-reviewed decision support for telemedicine/triage environments. However, the limited sample size and absence of external multi-center and prospective testing by clinicians with the weathered data limits clinical generalizability. Thus, the proposed system is not intended as a diagnostic or a prescription tool for treatment, but as a decision support system in a research phase.

Received: 24 April 2026 | Revised: 27 July 2026 | Accepted: 15 September 2026 

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

We confirm that the image datasets described in this study are publicly available in the following repositories: 1. Pharyngitis Images Dataset Kaggle: https://www.kaggle.com/datasets/bebeaa/pharyngitis-images-dataset. 2. PGUPharyngitis Figshare: https://doi.org/10.6084/m9.figshare.28163513.

Author Contribution Statement

Mizanur Rahman: Conceptualization, Methodology, Software, Formal analysis, Data curation, Writing – original draft. Fakrul Alam: Validation, Formal analysis, Data curation. Azharul Islam: Methodology, Validation, Formal analysis. Mazharul Islam: Data curation, Writing – review & editing. Sumon Rana: Formal analysis, Writing – review & editing. Firoz Kabir: Validation, Formal analysis, Writing – review & editing. Irfan Sadiq Rahat: Software, Writing – original draft, Supervision.