A research team has introduced ClinFusion-Net, a hybrid deep learning framework designed to achieve accurate and automated classification of gastrointestinal diseases from endoscopic images. By combining convolutional neural networks and vision transformers, the proposed system demonstrates high diagnostic classification performance alongside dedicated mechanisms for model explainability.
The design of ClinFusion-Net addresses endoscopic image interpretation by uniting two complementary deep learning paradigms into a unified pipeline. To process complex endoscopic visuals, the architecture utilizes ConvNeXt-XL to extract fine-grained local visual features from mucosal surfaces. Operating in parallel, Swin-Large captures broader contextual information across the entire endoscopic scene.
To integrate the two feature streams effectively, the framework incorporates a specialized MedicalAttention mechanism guided by class prototypes.
The training procedure for ClinFusion-Net relies on a composite objective function termed MedicalLoss, formulated to overcome typical diagnostic classification challenges. MedicalLoss combines three distinct loss components: weighted cross-entropy, focal loss, and attention-entropy regularization. This integrated loss structure was formulated to improve the model’s sensitivity across different disease classes and to refine predictive calibration during optimization.
Empirical validation of ClinFusion-Net was conducted across two benchmarks: the original Kvasir v1 dataset and its expanded successor, Kvasir v2. The research team trained and evaluated the framework separately on each dataset using release-specific data partitions. On the expanded Kvasir v2 release, ClinFusion-Net achieved a classification accuracy of 95.1 percent and a macro-average area under the receiver operating characteristic curve (macro-AUC) of 0.998. When assessed on the Kvasir v1 dataset, the model attained an accuracy of 96.3 percent and a matching macro-AUC of 0.998.
Comparative analysis on Kvasir v2 showed that ClinFusion-Net’s 95.1 percent accuracy was comparable to that of the strongest single-backbone baseline, which reached 95.0 percent. However, the researchers emphasized that ClinFusion-Net maintained strong class-balanced discrimination across the diagnostic categories through its integrated local and global representation. Controlled ablation experiments confirmed that the fusion of hybrid features served as the principal contributor to overall performance. In addition, the ablations revealed that the class-prototype attention mechanism and the composite MedicalLoss objective offered complementary gains, specifically improving model calibration and classification sensitivity on difficult-to-distinguish classes.
Beyond raw diagnostic metrics, the study placed substantial emphasis on interpretability to verify that model predictions align with clinically meaningful features. The authors implemented three complementary interpretability techniques: Grad-CAM, LIME, and t-SNE embedding visualizations.
The investigators noted specific methodological boundaries regarding the interpretation of their experimental findings. Because Kvasir v1 and Kvasir v2 represent related releases originating from the same underlying dataset family, the authors explicitly presented their findings as separate within-release evaluations instead of claiming evidence of cross-dataset generalization. Within this defined experimental scope, the study concludes that ClinFusion-Net provides an accurate and interpretable solution for automated gastrointestinal disease classification, holding strong potential to support clinical decision-making and facilitate earlier disease detection in real-world environments.
Subject of Research: Automated gastrointestinal lesion classification in endoscopic imaging
Article Title: Explainable GI lesion classification via class prototype guided hybrid feature fusion on Kvasir datasets
Article References: Habiby, A. T., Bushra, A., Assaduzzaman, M., Moon, M. M. H., Fahad, N., Mahmud, A., Liew, T. H., & Ohidujjaman (2026). Explainable GI lesion classification via class prototype guided hybrid feature fusion on Kvasir datasets. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09781-w
Image Credits: AI Generated
DOI: 10.1007/s42452-026-09781-w
Keywords: Gastrointestinal Disease, Endoscopy, Deep Learning, ClinFusion-Net, Medical Image Analysis, Vision Transformers, Explainable Artificial Intelligence, Kvasir Dataset, Explainable, lesion, classification, class
News Source: Blake Davidson. (November 9, 2026). Explainable GI lesion classification via class prototype guided hybrid feature fusion on Kvasir datasets. Scienmag.



