PolyNet: A self-attention based CNN model for classifying the colon polyp from colonoscopy image

Abstract

Colon polyps are small, precancerous growths in the colon that can indicate colorectal cancer (CRC), a disease that has a significant impact on public health. A colonoscopy is a medical procedure that helps detect colon polyps. However, the manual examination for identifying the type of polyps can be time-consuming, tedious, and prone to human error. Automatic classification of polyps through colonoscopy images can be more efficient. However, there are currently no specialized methods for the classification of polyps from colonoscopy; however, several state-of-the-art CNN models can classify polyps. We are introducing a new CNN-based model called PolyNet, a model that shows the best accuracy of the polyps classification from the multiple models and which also performs better than pre-trained models such as VGG16, ResNet50, DenseNetV3, MobileNetV3, and InceptionV3, as well as nine other customized CNN-based models for classification. This study provides a sensitivity analysis to demonstrate how slight modifications in the network's architecture can impact the balance between accuracy and performance. We examined different CNN architectures and developed a good convolutional neural network (CNN) model for correctly predicting colon polyps using the Kvasir dataset. The self-attention mechanism is incorporated in the best CNN model, i.e., PolypNet, to ensure better accuracy. To compare, DenseNetV3, MobileNet-V3, Inception-V3, VGG16, and ResNet50 get 73.87 %, 69.38 %, 61.12 %, 84.00 %, and 86.12 % of accuracy on the Kvasir dataset, while PolypNet with attention archives 86 % accuracy, 86 % precision, 85 % recall, and an 86 % F1-score.