A hybrid ViT-L/32-MaxViT-L architecture with adaptive gated fusion for multiclass gastrointestinal disease detection and multi-method post-hoc explainability.

INTRODUCTION: Accurate detection of gastrointestinal diseases from endoscopic images remains challenging due to substantial inter-class similarity, intra-class variability, and heterogeneous lesion presentation. While convolutional neural networks (CNNs) effectively capture fine mucosal textures, their limited receptive field restricts global contextual reasoning. Conversely, transformer architectures model long-range dependencies but may underrepresent localized structural detail.

METHODS: This study proposes a dual-branch hybrid transformer framework that integrates global token-level reasoning from ViT-L/32 with hierarchical spatial modeling from MaxViT-L. An adaptive gated fusion mechanism is introduced to dynamically regulate the relative contribution of local and global representations on a per-sample basis, enabling content-aware cross-scale integration. The model was evaluated on a multi-class gastrointestinal endoscopic dataset using comprehensive quantitative metrics, class-wise analysis, ablation studies, confidence interval estimation, and statistical validation. Post-hoc explainability was further investigated using multiple complementary XAI techniques.

RESULTS: The proposed framework demonstrated consistently high accuracy, sensitivity, specificity, and area under the ROC curve across eight gastrointestinal disease categories, outperforming several state-of-the-art transformer and hybrid architectures. The XAI analyses further confirmed lesion-focused activation consistency and stable decision-making.

DISCUSSION: The findings indicate that adaptive local-global feature integration through gated fusion enhances both diagnostic performance and interpretability, supporting the potential of hybrid transformer architectures for reliable computer-aided gastrointestinal disease classification.