PolypNet: An Attention Based Shallow Deep Learning Model for Colorectal Polyp Image Classification

Colon polyps are small, pre-cancerous growths in the colon that can indicate the presence of colorectal cancer (CRC), a disease that significantly impacts the well-being of individuals globally. Colonoscopy is a medical diagnostic technique used to find colon polyps. However, the manual examination and classification of polyps can be time-consuming, repetitive, and prone to human errors. Automation can improve the effectiveness of polyp classification in colonoscopy images. More specialized methodologies are needed to classify polyps identified after colonoscopy. Moreover, in the small dataset the deep learning model shows the overfitting problem. In this regard, this research presents a PolypNet, a shallow CNN-based deep-learning (DL) method combined with a self-attention mechanism for classifying colorectal polyps automatically and overcoming the overfitting problem. To evaluate the performance of our model, we do several assessments using the highly acknowledged open-source benchmark dataset Kvasir-v1, which has a total of 4000 images categorized into eight classes. The study compares four CNN based transfer-learning models for polyp classification, including VGG16, ResNet50, DenseNet-121, and MobileNet-v3. The proposed model achieves the accuracy of 0.86, nearly identical to ResNet50 while outperforming VGG16, DenseNet-121, and MobileNet-v3. The ablation study shows the significance of the proposed model.