Work

An Efficient Adversarial Self-Supervised Representation Learning Model for Classification of Anomalies in Wireless Capsule Endoscopy Images

Year: 2023
Type: dissertation
Abstract: Wireless capsule endoscopy (WCE) contains a tiny camera, a wireless transmitter, and a light. When a patient swallows a capsule, it travels down the human digestive system passing through the stomach, intestines, colon, and rectum, and the camera wirelessly captures images that are transmitted to an external recording device. Self-supervised learning (SSL) consists of a pretext task, where the network learns data representations from a large-scale unlabeled dataset while doing a task, followed by a downstream task, where the network uses the knowledge learned in the pretext task to do a downstream task with a small-scale dataset. SSL has transformed artificial intelligence (AI) by enhancing generalizability and reducing the reliance on vast amounts of labeled data. However, the application of SSL models in the field of WCE has received insufficient attention. The existing SSL models demand extensive data and computational resources during the pre-training phase. However, in the WCE field, only limited number of small-sized datasets are accessible which are from different domains and collected using different instruments. Prior research in SSL models for WCE relies on custom datasets exclusively available to specific studies. This research introduces an efficient SSL model that overcomes data scarcity and computational limitations and shows a high-level domain-free quality on a dataset captured by a different instrumentation. Our model is able to achieve 92% accuracy in a 4-category classification task and 74% on an 8-category classification task by employing a small-scale GAN to pre-train on a significantly smaller dataset (for example, about 170 times smaller than the ImageNet dataset used for pre-training the well-known SimCLR model proposed by Google), called the Kvasir dataset (The biggest public WCE dataset containing 8000 images from 8 different categories), compared to current pioneering SSL models. Subsequently, a small multi-layer perceptron (MLP) classifier is trained on the learned representations from the GAN's discriminator network for the classification task, employing a small-scale, labeled dataset from a different set of the Kvasir dataset. To ensure computational efficiency, we employ a compact GAN architecture with only two convolutional hidden layers in both the generator and discriminator networks. To overcome challenges with generating high quality synthetic data using a tiny GAN network and a small-scale dataset, we propose a reverse-order architecture for the GAN, where the discriminator and generator share the same architecture but in reverse order. Based on experiments, the GAN network based on a reverse-order architecture acquires better results in generating high quality samples using a small-scale dataset and small-size networks. By developing this novel adversarial SSL model for low data and computational resource scenarios, we address the limitations of existing models in terms of data supply and computational resources in the WCE field, as our model has acquired the same level of performance as its supervised counterpart trained on a much bigger labeled dataset. Also, we have pre-trained a SimCLR model using the same data and computational resources as that of the proposed adversarial model, and the proposed adversarial SSL model has outperformed the SimCLR in every classification task. The proposed adversarial model has shown a robust performance against domain change as we trained the MLP classifier on the KID dataset, which is captured by a different instrumentation than that of the Kvasir dataset, on top of the representations learned from the Kvasir dataset. The proposed model obtained a high level of accuracy on the KID dataset illustrating that the model is robust against domain change.
Author: Jamali, Ali
Language: English
Open Access status: green
OpenAlex ID: W6990622514
Cites: 52
Cited by: 1
Related to: 9