Enhanced Ulcer Detection Using GAN-Augmented Image Processing and Convolutional Neural Networks

Abstract

Ulcers, especially chronic wounds and pressure sores, create significant healthcare challenges and need early detection for effective treatment. Due to the persistent challenges posed by human error in manual detection and classification processes, there is a need for automated methodologies that can offer robust and reliable solutions. To enhance the accuracy and generalization of the Convolutional Neural Network (CNN), a Generative Adversarial Network (GAN) based image augmentation technique is used in this study to classify the Kvasir dataset. With an initial 1512 ulcer images selected from the dataset, after stable 70 epochs, 70000 images were generated using GAN for 2 classes: Ulcer or Non-Ulcer. To remove the noise from an enhanced batch of images and classify the set of newly generated datasets, GAN-based CNN (G-CNN) was employed to get 99.00% training and 96.04% validation accuracy. The study was compared with conventional CNN which achieved a training accuracy of 98.5% and validation accuracy of 94.34%. A significant improvement has been observed in f1-scores of 0.94 and 0.97 for CNN and the proposed algorithm, respectively. The study is replicable across datasets with a limited number of images, thereby facilitating heightened accuracy levels.