Researchers have introduced a dual-branch deep learning architecture named SeAttnFusionNet, designed to automate the binary classification of pharyngitis versus no pharyngitis from standard RGB throat images. The study, published in Discover Analytics in June 2026, addresses the inherent subjectivity and variability in visual throat assessments by providing a consistent, repeatable computational method for screening. By leveraging complementary feature extraction pathways, the model aims to improve the reliability of automated decision support in primary care settings where rapid triage is essential.
The central challenge in automated throat image analysis lies in the substantial intra-class variability and the strong visual similarity between mild inflammation and healthy appearances. Factors such as illumination changes, viewpoint shifts, and acquisition artifacts like saliva glare or motion blur often obscure the subtle visual cues associated with pharyngitis. Conventional single-branch convolutional models frequently struggle with these conditions, as they may overfit to spurious background cues or dominant textures rather than the underlying pathology. SeAttnFusionNet was developed to mitigate these issues by explicitly separating feature learning into two distinct but synergistic pathways.
The architecture comprises a texture-focused branch and a context-focused branch. The texture branch utilizes depthwise-separable convolutions and stacked Squeeze-and-Excitation blocks to refine fine-grained cues such as edges, micro-textures, and localized intensity transitions. These features are critical for detecting subtle inflammatory changes in tissue coloration and surface texture. In parallel, the context branch employs multi-scale convolutions and local window attention mechanisms to capture broader structural patterns, including tonsillar symmetry and region-level redness distribution. This dual approach ensures that the model does not prematurely discard localized evidence or bias toward coarse appearance statistics.
A key methodological innovation in SeAttnFusionNet is the feature-map level fusion that occurs prior to global average pooling. By concatenating the outputs of the two branches at the feature-map stage, the model preserves spatially resolved complementary representations. This strategy allows the network to integrate local and contextual descriptors without losing location information, which is essential for accurate discrimination in small, discriminative regions. The fused representation is then passed through a compact classification head consisting of dense layers, dropout for regularization, and a sigmoid output layer for probability estimation.
The evaluation protocol employed a leakage-safe 5-fold cross-validation with a per-fold 80/20 split on a dataset of 312 RGB throat images. The dataset consisted of 185 images labeled as no pharyngitis and 127 images labeled as pharyngitis, indicating a moderate class imbalance. In each fold, the training set included 250 images, while the held-out test fold contained 62 images. This strict isolation of test data during training iterations was designed to prevent information leakage and provide a conservative estimate of model generalization performance under realistic variability.
Under this rigorous evaluation framework, SeAttnFusionNet achieved a cross-validation validation accuracy of 97.26%, with a precision of 97.58%, recall of 97.44%, and an F1-score of 97.51%. The confusion matrix revealed 168 true negatives, 17 false positives, 3 false negatives, and 124 true positives. These results indicate a balanced operating profile, with the model maintaining high sensitivity to pharyngitis cases while keeping false alarms at a moderate level. The close alignment between training and testing metrics suggests that the model does not rely on fold-specific artifacts and maintains stable performance when exposed to unseen data.
Comparative analysis against baseline models further highlighted the efficacy of the proposed architecture. ConvNeXt recorded a CV validation accuracy of 97.02%, while CoAtNet achieved 95.86%, Swin Transformer reached 94.17%, and NASNetMobile yielded 89.71%. SeAttnFusionNet outperformed all baselines in both accuracy and F1-score, demonstrating that the combination of multi-scale attention-driven context modeling with efficient texture refinement provides a practical and reliable framework for automated pharyngitis classification. The ablation study confirmed that both single-branch variants performed worse than the full model, supporting the hypothesis that pharyngitis recognition benefits from the joint use of fine-grained texture cues and broader contextual information.
To address the interpretability concerns inherent in deep learning models, the researchers employed Gradient-weighted Class Activation Mapping, or Grad-CAM, to visualize the spatial regions contributing most strongly to the model’s predictions. The visualizations showed that SeAttnFusionNet concentrated its activation more consistently over anatomically plausible inflammatory regions, such as the tonsillar zones and adjacent pharyngeal tissue. In contrast, baseline models like CoAtNet and ConvNeXt exhibited scattered responses that extended into anatomically irrelevant regions. This qualitative evidence supports the notion that the proposed model aligns its predictions with relevant throat structures, enhancing transparency without claiming clinical validation.
Despite the strong quantitative results, the study acknowledges several limitations that constrain the interpretation of the findings. The dataset size of 312 images is relatively small for training a dual-branch deep learning architecture, which limits statistical robustness and increases the possibility that performance is influenced by dataset-specific characteristics. Furthermore, the study did not include evaluation on an independent external dataset, meaning the reported metrics reflect performance only on images drawn from the same underlying distribution. External validation on data collected from different devices, environments, and patient groups is identified as a critical next step for establishing broad generalizability.
The clinical framing of the task is also limited to a binary classification of pharyngitis versus no pharyngitis, which does not fully represent the primary clinical decision problem in routine practice. In real-world settings, the more consequential distinction is often between bacterial and viral pharyngitis, as this differentiation directly impacts treatment selection and antibiotic stewardship. RGB throat images alone may not sufficiently distinguish between these etiologies due to considerable visual overlap. Therefore, the current findings should be viewed as an encouraging proof of concept for screening-oriented image classification rather than a complete etiologic diagnostic solution.
Future work will prioritize larger and more diverse multi-center data collection to improve statistical power and reduce sensitivity to site-specific imaging characteristics. Researchers plan to conduct external validation on independent datasets and investigate robustness to acquisition variability, including differences in lighting, device characteristics, and patient posture. Additionally, the framework may be extended toward multi-class or multimodal settings, integrating symptom profiles and laboratory information to better align the modeling objective with real-world diagnostic workflows. These steps are necessary to establish the reliability and suitability of the model for real-world adoption.
The development of SeAttnFusionNet represents a significant step toward more reliable and transparent automated pharyngitis image analysis. By combining architectural innovations with rigorous leakage-safe evaluation and interpretability analysis, the study provides a robust foundation for future research in medical image classification. The emphasis on preserving spatial evidence and ensuring methodological integrity addresses key challenges in small medical image datasets, offering a template for developing trustworthy AI tools in clinical decision support.
Subject of Research: SeAttnFusionNet for pharyngitis versus no pharyngitis classification from RGB throat images with grad-CAM-based interpretability
Article Title: SeAttnFusionNet for pharyngitis versus no pharyngitis classification from RGB throat images with grad-CAM-based interpretability
Article References: Meghla, H. M., Rahim, A., Shehan, M. S. H., Melon, M. M. H., Shahiduzzaman, M., Alam, M. F., Rahman, M. M., & Tesfaye, A. (2026). SeAttnFusionNet for pharyngitis versus no pharyngitis classification from RGB throat images with grad-CAM-based interpretability. Discover Analytics, 4(1), Article 22. https://doi.org/10.1007/s44257-026-00073-6
Image Credits: AI Generated
DOI: 10.1007/s44257-026-00073-6
Keywords: SeAttnFusionNet, pharyngitis, versus, classification, throat, images, grad-CAM-based, interpretability, scientific research
News Source: Denise Maddox. (November 8, 2026). Dual-Branch AI Model Classifies Throat Inflammation from Images. Scienmag.



