Autoencoder- Based Multistage Strategy for Class Imbalance in Medical Imaging Analysis - Chest X-rays

Academics and professionals in the industry are interested in the challenge of unbalanced learning. When one class significantly outnumbers the others, a class imbalance issue arises, resulting in most classical classifiers performing poorly on the minority classes. Numerous strategies have been proposed to address this issue, with most of them revolving around resampling techniques. In this work, the objective is to determine the optimal model architecture that uses a multistage approach to address the issue of class imbalance in medical imaging utilizing Convolutional Neural Network (CNN)-based Autoencoder methodologies. The proposed approach addresses the problem in multiple stages: Stage one is training an autoencoder that discriminates both the classes with reconstruction of images, stage two is extracting features from the trained encoder section for both the classes and stage three is building an Artificial neural network for classifying the disease. The experiments produced the results of 100% on train data and 90.90% overall recall on minority samples while 97.06% accuracy on test data.