DFPT-CNN: A Dual Feature Extraction and Pretrained CNN Synergy for Minimal Computational Overhead and Enhanced Accuracy in Multi-Class Medical Image Classification

Abstract

In the advanced computer vision era, Convolutional Neural Network (CNN) plays a pivotal role in image processing, as they excel at automatically extracting important patterns, and structures, for accurate analysis across diverse domains. However, achieving higher accuracy often leads to intensifying computational and timing demands. To address the challenge, this research introduces a novel dual feature extraction methodology. This approach is implemented using two distinct feature extraction modules, employed at different stages of the model: (1) Edge Gradient-Dimensionality Reduction (EGDR) module which encapsulates the extraction of pixel edge gradient features from the raw input frame, leading to a dimensionality reduction by a factor of 0.5; (2) Subtle Local Feature Extraction (SLFE) pooling algorithm module, prioritizes the extraction of local and subtle features over maximum or average feature content. The combination of these two stages proves particularly effective in enhancing classification accuracy while minimizing computational overhead and training duration. Subsequently, comprehensive training, validation, and testing were conducted on a selected multi-class chest computed tomography medical image dataset using various state-of-the-art CNN architectures such as VGG-16, InceptionV3, ResNet50 to identify the most suitable model for further experimentation with the proposed method. The proposed CNN-SLFE framework with EGDR module achieved a significant reduction of 17.94% in computational time compared to non-EGDR module, and concurrently enhanced the classification accuracy with an improvement factor of 1.17 compared to existing CNN frameworks with EGDR module.