Quaternion Image Enhancements with Classifier Ensembles for Gastrointestinal Disorder Detection and Resiliency to Contrast Degradation Adversarial Attack

Gastrointestinal (GI) diseases if undiagnosed can lead to severe health complications, including malnutrition, dehydration, and even cancer. Colonoscopy, the gold standard for GI disorder diagnosis, often faces challenges such as low-contrast images due to poor illumination, making it difficult to detect fine details crucial for accurate diagnosis. Additionally, AI-based healthcare, including GI Disease Detection systems, is susceptible to adversarial attacks, raising significant concerns for patient safety due to potential misclassifications in clinical decision-making. To mitigate these issues, we introduce a novel Computer Assisted Diagnostics (CAD) system for GI disease detection and resilience against adversarial threats. Our approach incorporates two key innovations: first, a preprocessing technique using Quaternion Dark Channel Laplacian of Gaussian (QDC-LoG) for edge and contrast improvements; second, an ensemble of Convolutional Neural Networks (CNNs) and classifiers for robust GI disorder prediction. Through comprehensive experiments, we demonstrate that our method not only surpasses current state-of-the-art techniques, including advanced deep learning models and transformers, by achieving a 97.88% accuracy on the Wireless Capsule Endoscopy (WCE) Curated Colon Disease Dataset but also shows exceptional specificities on Kvasir V1 and V2 datasets. Furthermore, our system exhibits strong resilience against Contrast Adversarial Degradation Attacks (CADA), ensuring reliable performance even under severe contrast reduction.