A lightweight hybrid CNN framework with Multi-filter feature selection and Selective Opposition-based Grey Wolf Optimization for gastrointestinal image classification

Abstract Gastrointestinal disease detection through endoscopic imaging has become a critical research area in computer-aided medical diagnosis. Accurate classification of endoscopic images remains challenging due to high intra-class variability and visual similarity across pathological categories. In this work, an efficient deep feature extraction and optimization-based framework for automated gastrointestinal image classification using the Kvasir v1 and Kvasir v2 datasets is proposed. Initially, multiple pretrained convolutional neural network models were evaluated as baseline feature extractors. To achieve a balance between classification performance and computational efficiency, an ensemble of MobileNetV2, EfficientNet-B0, and ConvNeXt-Tiny was employed to extract complementary deep features from endoscopic images. The fused feature representation achieved an initial classification accuracy of 96.12% on Kvasir v1 and 95.43% on Kvasir v2. To reduce feature redundancy and enhance discriminative capability, a hybrid filter-based feature selection strategy combining Hilbert-Schmidt Independence Criterion, Laplacian Score, and Elastic Net was applied. The selected features were further refined using a novel Selective Opposition-based Grey Wolf Optimizer, which improved exploration capability and enabled effective identification of the most informative feature subset. The optimized features were finally classified using a multilayer perceptron classifier. Experimental results show that the proposed framework achieved superior classification accuracies of 97.51% on Kvasir v1 and 97.06% on Kvasir v2, outperforming baseline CNN models and conventional feature fusion approaches. These results demonstrate that the integration of lightweight deep feature extractors, hybrid feature selection, and metaheuristic optimization provides a robust and computationally efficient solution for automated gastrointestinal image classification across multiple benchmark datasets.