Objective:

To design a dual-stream transformer architecture integrating ViT-L/32 and MaxViT-L for capturing long-range contextual dependencies and fine-grained mucosal structures, and to develop an adaptive gated fusion mechanism for balancing global and local feature representations.

Approach:

• Model Framework: A dual-branch hybrid transformer framework integrating global token-level reasoning from ViT-L/32 with hierarchical spatial modeling from MaxViT-L.
• Adaptive Gated Fusion: An adaptive gated fusion mechanism regulates the contribution of local and global representations on a per-sample basis.
• Evaluation: The model was evaluated on a multi-class gastrointestinal endoscopic dataset using quantitative metrics, class-wise analysis, ablation studies, confidence interval estimation, and statistical validation.
• Explainability: Post-hoc explainability was investigated using multiple complementary XAI techniques.

Key Findings:

• The proposed framework achieved high accuracy, sensitivity, specificity, and area under the ROC curve across eight gastrointestinal disease categories.
• It outperformed several state-of-the-art transformer and hybrid architectures.
• XAI analyses confirmed lesion-focused activation consistency.

Interpretation:

Adaptive local–global feature integration through gated fusion enhances diagnostic performance and interpretability.

Limitations:

• The complexity of deep architectures raises concerns regarding interpretability.
• Many transformer architectures do not integrate global and local modeling effectively.

Conclusion:

The study supports the potential of hybrid transformer architectures for reliable computer-aided gastrointestinal disease classification.