<p>Gastrointestinal illnesses (GIDs) provide a substantial global public health burden, necessitating early and precise diagnosis to enhance patient outcomes. Nevertheless, owing to the intricate backdrops and varied morphological traits of lesions in endoscopic pictures, conventional approaches are susceptible to misdiagnoses and overlooked detections. This study introduces an Adaptive Dual-Perception Contrastive Learning Framework (ADP-CLF) to tackle these issues in the classification of gastrointestinal disorders, with the objective of improving diagnostic efficiency and accuracy. By combining fine-grained feature modelling in the local pathway with nearest-neighbor contrastive learning in the global pathway, ADP-CLF makes it possible to precisely capture important information on several scales. The system includes a perception module that utilises deformable convolutions to improve the modelling of irregular lesion shapes, augmented by a self-attention mechanism to better the depiction of critical areas. The framework enhances class consistency in the feature space by a nearest-neighbor positive sample creation technique, thereby markedly increasing classification performance and out-of-distribution detection. Experiments on the Kvasir-dataset-v2 and Kvasir-capsule public datasets indicate that the suggested method surpasses state-of-the-art techniques in classification accuracy, precision, recall, and F1-score, with a maximum classification accuracy of 98.66%. These findings confirm the outstanding efficacy of ADP-CLF in intricate medical picture processing, offering a solid basis for intelligent gastrointestinal disease detection and precision medicine.</p>