Gastro Intestinal Disease Classification Using Hierarchical Spatio Pyramid TranfoNet With PitTree Fusion and Efficient-CondConv SwishNet

Early detection of Gastrointestinal (GI) tract diseases is essential for effective healthcare management, treatment, and prevention, ultimately lowering morbidity and mortality rates worldwide. Current classification models lack spatial feature arrangement consideration, diminishing discriminative power and leading to misdiagnosis of esophagitis and ulcerative colitis due to overlapping visual characteristics with other GI diseases. Hence, a novel Hierarchical Spatio Pyramid TranfoNet featuring a Spatial Transformer Network (STN) with spatial pyramid pooling is introduced, which enhances discriminative power in distinguishing between overlapping disease characteristics. Enhancing classification models for Dyed Lifted Polyps (DLP) and Dyed Resection Margins (DRM) in endoscopy images is critical for precise gastrointestinal diagnosis, tackling challenges posed by spatial complexity and inter-class confounders. Hence, a novel PitTree Fusion Algorithm, combining Minimum Spanning Tree (MST) analysis and Kudo's pit pattern analysis is introduced to accurately locate and differentiate normal tissue from dyed regions like DLPs and DRMs in endoscopy images. Then, a novel Efficient-CondConv SwishNet is introduced to enhance GI disease classification by extracting informative features from endoscopic images, utilizing EfficientNet-CondConv with Swish activation. After classification, heatmaps highlighting influential regions are produced via gradient-weighted class activation mapping, or Grad-CAM, which provides information about classification decisions. The proposed Hierarchical Spatio Pyramid TranfoNet with PitTree Fusion and EfficientNet-CondConv SwishNet achieved a classification accuracy of 98.2%. The proposed framework is tested on 8000 images using the Kvasir dataset, which is publicly available in Kaggle, consisting of eight classes. The results show that the suggested model outperforms the current models showing increased accuracy, precision, recall, sensitivity, specificity, F1 score, and reduced loss rate.