Bottleneck Feature-Based U-Net for Automated Detection and Segmentation of Gastrointestinal Tract Tumors from CT Scans

Abstract

In today's medical landscape, an array of diagnostic techniques for cancer, leveraging imaging data, have become increasingly prevalent. This has posed a unique challenge for radiologists in the detection of Digestive System Cancer (DSC). This paper introduces the Bottleneck Feature-based U-Net, an innovative method designed for the automated detection and segmentation of the digestive system utilizing endoscopy. The U-Net design, previously proven successful for image segmentation tasks, is harnessed to its full potential in our proposed method. We have enhanced its performance by integrating a bottleneck feature extraction technique. The encoding U-Net is initially trained prior to the training of the BS U-Net, facilitating the procurement of encodings from label maps containing crucial anatomical information, such as shape and location. A Bottleneck Supervised (BS) U-Net is thus formed by pairing an encoding U-Net with a segmentation U-Net. Our proposed bottleneck feature in the U-Network enables the model to compress input data, an essential learning component. This compressed view of data retains vital information used for either reconstructing the input image or carrying out the segmentation process. In the current study, we put forth a bottleneck-based U-Net model tailored to perform gastrointestinal tract tumor segmentation. To train and test our method, we employed the comprehensive Kvasir dataset, which encompasses a wide range of digestive system images. We further tested the robustness and generalizability of our model through a thorough quantitative and qualitative analysis. The results underscore the versatility of the bottleneck U-Net and its potential as a reliable tool for radiologists in clinical practice. Our proposed model demonstrated rapid and effective cancer diagnosis capabilities, thus reducing diagnosis time. The model exhibited an impressive accuracy rate of 98.64% and a specificity score of 99.71%, outperforming both LSTM-ANN and GA Algorithms. This not only attests to the efficacy of our model but also underscores its potential in advancing diagnostic methodologies in clinical settings.