EndoNet: A Multiscale Deep Learning Framework for Multiple Gastrointestinal Disease Classification via Endoscopic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Gastrointestinal (GI) disorders pose significant healthcare challenges, necessitating advanced diagnostic tools.
• Wireless capsule endoscopy (WCE) aids GI abnormality detection, but differentiating similar lesions is difficult.
• Existing computer-aided diagnostic (CAD) systems often lack the complexity to analyze diverse GI disease characteristics.

Purpose of the Study:

• To develop and evaluate EndoNet, a multi-stage hybrid deep learning (DL) framework for classifying eight types of GI diseases using WCE images.
• To improve the accuracy and interpretability of automated GI disease diagnosis.
• To provide a generalizable AI solution for clinical decision support.

Main Methods:

• Proposed EndoNet, a hybrid DL framework integrating features from multiple pre-trained convolutional neural networks (CNNs) (Inception, Xception, ResNet101).
• Employed inter-layer and inter-model feature fusion, Non-Negative Matrix Factorization (NNMF) for dimensionality reduction, and Minimum Redundancy Maximum Relevance (mRMR) for feature selection.
• Evaluated performance on the Kvasir v2 and HyperKvasir datasets using seven machine learning algorithms.

Main Results:

• EndoNet achieved high classification accuracy, reaching up to 97.8% on the Kvasir v2 dataset and 98.4% on the HyperKvasir dataset.
• The framework demonstrated effective feature extraction, fusion, and selection for complex GI image analysis.
• The combination of DL and feature engineering proved superior to traditional CAD approaches.

Conclusions:

• EndoNet integrates transfer learning, feature engineering, dimensionality reduction, and feature selection for accurate GI disease classification.
• The proposed framework offers high accuracy, flexibility, and interpretability, making it suitable for clinical decision support.
• EndoNet represents a powerful and generalizable AI solution for advancing WCE-based gastrointestinal diagnostics.