A Novel CSRU-GAN Framework for Polyp Segmentation

Polyp segmentation plays a crucial role in the early detection and diagnosis of fatal diseases such as colorectal cancer during colonoscopy procedures. Timely diagnosis of such catastrophic diseases could assist medical experts and potentially avert impending crisis Over the years, image segmentation in biomedical applications has been undergoing a massive surge especially with the introduction of the U-Net architecture in 2015 and the FCN (Fully Convolutional Network). This architecture went on to become a traditional thresholding and edge-detection technique. Several years hence, more sophisticated deep learning models of the U-Net such as Attention U-Net and ResU-Net were introduced. These architectures have significantly enhanced segmentation accuracy by capturing complex spatial hierarchies. However, conventional encoder-decoder frameworks still face limitations in capturing long-range dependencies and generating fine-grained boundaries in challenging scenarios. In order to tackle the aforementioned issues, several GAN (Generative Adversarial Network) models were introduced to enhance the quality of segmentation masks. These architectures massively outperformed the conventional architectures in terms of dice and IOU scores which are touted to be the primordial metrics designed to evaluate the authenticity of the masks generated by the models. This paper attempts to propose a new Conditional Generative Adversarial Network model called CSRU-GAN (Conditional SWIN ResU-Net Generative Adversarial Network) which intends to use the ability of the residual connections inspired from the ResNet architecture to clip the vanishing gradient problem, the ability of SWIN transformers to capture hierarchical features by making use of shifted windows and attention mechanism. Lastly, it makes use of the adversarial training of the CGAN (Conditional Generative Adversarial Network) while also taking into account the ground truth information. The proposed model when tested on the CVC database performed admirably in terms of the Dice and IOU (Intersection Over Union) scores which were found out to be $\mathbf{9 1. 2 1 \%}$ and $\mathbf{8 3. 2 9 \%}$ respectively.