Problem

Medical image classification, particularly for conditions like COVID-19, frequently encounters the significant challenge of imbalanced datasets. This issue arises when there is a stark disparity in the number of images across different classes, such as having many normal chest X-rays but very few COVID-19 positive cases. Such data scarcity and imbalance can severely impede the performance and reliability of artificial intelligence and machine learning algorithms, making accurate and rapid disease detection difficult, especially during public health crises.

Method

This study introduces a novel approach to tackle imbalanced medical image data by integrating a Progressive Generative Adversarial Network (ProGAN) with a Slime Mould Algorithm (SMA)-optimized ResNet50V2 classifier. The method involves several key steps. First, a customized ProGAN is individually trained for each classification class (COVID-19, Normal, Viral Pneumonia, and Lung Opacity) to generate high-quality synthetic chest X-ray images, aiming to address data scarcity. Second, a weighted synthetic image injection strategy is employed to combine these generated images with the real dataset. This weighting ensures that classes with fewer real images receive a proportionally larger number of synthetic images, thereby balancing the dataset. Finally, a pre-trained ResNet50V2 deep learning model is used as the classifier, with its critical hyperparameters (learning rate, synthetic image injection ratio, and dropout rate) optimized using the Slime Mould Algorithm (SMA), a population-based meta-heuristic optimization technique. The overall framework is designed to improve classifier performance on imbalanced medical datasets.

Results

The proposed model demonstrated significant improvements in medical image classification on an imbalanced chest X-ray dataset. For a 4-class classification problem (Normal, Viral Pneumonia, COVID-19, Lung Opacity), the model achieved an accuracy of 95.5%. When applied to a 2-class classification problem, the accuracy further increased to 98.5%. The optimization process identified an optimal synthetic image injection rate of 20% and a learning rate of 7.26e-5 for the ResNet50V2 network. The integration of ProGAN-generated synthetic images and SMA-optimized hyperparameters led to an enhancement in classification accuracy by 3.53% for the ResNet50V2 classifier. Furthermore, the synthetic images generated by the ProGAN, particularly in its later progressive stages, were qualitatively high, with many being indistinguishable from real images to a non-expert eye.

Implications

The successful outcomes of this study underscore the effectiveness of combining generative adversarial networks for data augmentation with meta-heuristic optimization for classifier tuning, particularly for medical image analysis with imbalanced datasets. This approach offers a robust solution for enhancing the accuracy and reliability of disease detection systems, especially in scenarios like pandemics where rapid and precise diagnosis from limited or skewed data is crucial. The methodology presents a valuable framework that can be adapted to various medical imaging tasks facing similar challenges of data scarcity and class imbalance, potentially accelerating early detection and improving patient outcomes by enabling more reliable AI-driven diagnostic tools.