Visual Feature Refinement with MECNET for Gastrointestinal Cancer Classification

Abstract

Early detection and classification of gastrointestinal tract pathologies are crucial for better prognosis and reduced mortality rates, such as in colorectal cancer. In this paper, we introduce MECNET, a new hybrid deep learning framework for efficient classification of endoscopic images. The proposed framework integrates the feature refiner module with state-of-the-art CNN architectures such as VGG19, ResNet50, and EfficientNet for improved performance in image analysis and classification tasks. The feature refiner module successively applies grayscale, Gaussian, and LPQ filters to extract meaningful texture features, which play an important role in differentiating disease categories. Our proposed scheme has been tested on several available datasets, namely WCE, Kvasir, GastroVision, and SCPolyp including 13,000 images from four categories: normal colon, polyps, esophagus, and ulcerative conditions. The MECNET model attained an appreciable performance metric, outperforming state-of-the-art methods at accuracy and F1 scores of 97.4% and 97.34% on the WCE test set and 97.2% and 97.26% on the Kvasir test set, respectively. This proves that MECNET does not only excel in classification but also generalizes well across diverse datasets. The novelty of this work lies inincorporating a feature refiner module with established CNN architectures and utilizinga hybrid ensemble approach. This approach will provide a boost to the model's performance. The proposed framework addresses key challenges in medical image classification: improving feature extraction by making full use of advanced transfer learning techniques.