Curated Colon Disease Diagnosis Using Principal Component Analysis and Deep Learning with Integrated Gradients

Accurate and timely diagnosis of gastrointestinal (GI) diseases is crucial for effective treatment and improved patient outcomes. Wireless Capsule Endoscopy (WCE) has emerged as a valuable tool for visualizing the GI tract, but manual analysis of WCE images is time-consuming and prone to variability. This study introduces ECPIG, a novel framework leveraging deep learning for automated and efficient GI disease classification using WCE images. ECPIG incorporates preprocessing, transfer learning, fine-tuning, and dimensionality reduction via Principal Component Analysis (PCA) applied to the EfficientNetB3 architecture. We evaluated ECPIG’s performance on a comprehensive dataset of 6000 WCE images encompassing four common GI conditions: normal, ulcerative colitis, polyps, and esophagitis. Model performance was rigorously assessed using accuracy, recall, F1-score, and precision to provide a comprehensive understanding of its classification capabilities. Our results demonstrate that ECPIG achieves state-of-the-art accuracy, outperforming existing methods while significantly reducing training time through PCA. Furthermore, we employed Integrated Gradients explanation to enhance the transparency and interpretability of our deep learning model. This study highlights the potential of ECPIG as an effective and efficient tool for automated GI disease classification, paving the way for improved clinical decision support and patient care.