Integrating Deep Feature Extraction and Hybrid ResNet-DenseNet Model for Multi-Class Abnormality Detection in Endoscopic Images

Aman Sagar,Preeti Mehta,Monika Shrivastva,Suchi Kumari
2024-10-24
Abstract:This paper presents a deep learning framework for the multi-class classification of gastrointestinal abnormalities in Video Capsule Endoscopy (VCE) frames. The aim is to automate the identification of ten GI abnormality classes, including angioectasia, bleeding, and ulcers, thereby reducing the diagnostic burden on gastroenterologists. Utilizing an ensemble of DenseNet and ResNet architectures, the proposed model achieves an overall accuracy of 94\% across a well-structured dataset. Precision scores range from 0.56 for erythema to 1.00 for worms, with recall rates peaking at 98% for normal findings. This study emphasizes the importance of robust data preprocessing techniques, including normalization and augmentation, in enhancing model performance. The contributions of this work lie in developing an effective AI-driven tool that streamlines the diagnostic process in gastroenterology, ultimately improving patient care and clinical outcomes.
Computer Vision and Pattern Recognition,Machine Learning
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to automatically identify and classify multi - class gastrointestinal abnormalities in video capsule endoscopy (VCE) images, so as to reduce the diagnostic burden on gastroenterologists and improve the diagnostic efficiency. Specifically, the paper focuses on the following aspects: 1. **Multi - class Abnormality Detection**: The paper aims to develop a deep - learning framework that can automatically identify and classify ten different types of gastrointestinal abnormalities, including telangiectasia, bleeding, ulcers, etc. This helps to reduce the time and workload of gastroenterologists in manually examining a large number of VCE images. 2. **Improving Diagnostic Accuracy**: By combining two convolutional neural network architectures, DenseNet and ResNet, the paper proposes a hybrid model to improve the classification accuracy of different types of gastrointestinal abnormalities. The experimental results show that this model performs excellently on multiple performance indicators, with an overall accuracy rate of 94%. 3. **Meeting Data Challenges**: The amount of VCE image data is huge and there are changes in visual conditions (such as bubbles, debris, food residues, etc.), which pose challenges to the robustness and generalization ability of the model. For this reason, the paper introduces advanced data pre - processing techniques, including normalization and data augmentation (such as random horizontal flipping and rotation), to ensure that the model can maintain good performance under different conditions. 4. **Clinical Application Potential**: Through automated and intelligent image analysis, this research is expected to accelerate the diagnostic process, reduce human errors, and ultimately improve the treatment effect and clinical outcome of patients. In summary, the core objective of this paper is to use deep - learning technology to develop an efficient and accurate multi - class gastrointestinal abnormality detection tool to support medical professionals in making faster and more accurate diagnoses.