Multiclassification of Colorectal Polyps from Colonoscopy Images Using AI for Early Diagnosis.

Colorectal cancer (CRC) remains one of the leading causes of cancer-related mortality worldwide, emphasizing the critical need for the accurate classification of precancerous polyps. This research presents an extensive analysis of the multiclassification framework leveraging various deep learning (DL) architectures for the automated classification of colorectal polyps from colonoscopy images. The proposed methodology integrates real-time data for training and utilizes a publicly available dataset for testing, ensuring generalizability. The real-time images were cautiously annotated and verified by a panel of experts, including post-graduate medical doctors and gastroenterology specialists. The DL models were designed to categorize the preprocessed colonoscopy images into four clinically significant classes: hyperplastic, serrated, adenoma, and normal. A suite of state-of-the-art models, including VGG16, VGG19, ResNet50, DenseNet121, EfficientNetV2, InceptionNetV3, Vision Transformer (ViT), and the custom-developed CRP-ViT, were trained and rigorously evaluated for this task. Notably, the CRP-ViT model exhibited superior capability in capturing intricate features, achieving an impressive accuracy of 97.28% during training and 96.02% during validation with real-time images. Furthermore, the model demonstrated remarkable performance during testing on the public dataset, attaining an accuracy of 95.69%. To facilitate real-time interaction and clinical applicability, a user-friendly interface was developed using Gradio, allowing healthcare professionals to upload colonoscopy images and receive instant classification results. The CRP-ViT model effectively predicts and categorizes colonoscopy images into clinically relevant classes, aiding gastroenterologists in decision-making. This study highlights the potential of integrating AI-driven models into routine clinical practice to improve colorectal cancer screening outcomes and reduce diagnostic variability.