Evidence-Efficient Compressed Deep Ensemble Resnet for COVID-19 Detection

Abstract

The new coronavirus illness 2019 (COVID-19) put enormous strain on healthcare systems globally. Early diagnosis of COVID-19 is critical for controlling the spread of the COVID-19 pandemic and relieving pressure on health-care systems. Deep Learning (DL) models have a significant role in the COVID-19 detection by adopting the medical imaging techniques like Computerized Tomography (CT) and Chest X-ray (CXR). Many DL models have been developed to detect and diagnose the COVID-19 cases at earlier stages. Amongst, Hybrid Deep Transfer Learning (HDTL) model is developed using an Enhanced Convolution Restricted Boltzmann Machine (ECRBM) and ResNet model to improve the Covid-19 recognition tasks. But still, this model results in epistemic uncertainty which effects the performance of DL models employed for COVID-19 detection and diagnosis. On considering this, an effective DL framework is proposed to solve the challenges of epistemic uncertainty in DL models to provide a rapid detection of various diseases using CT and CXR images.  This model considers the uncertainty issue from the theory of evidence perspective. The softmax layer of ResNet models is modified with Dirichlet density parameters to represent learner predictions as a distribution over possible softmax outputs. The model has a specific loss function which is minimized through standard backprop and fitted to data by minimizing the Bayes risk with respect to the L2-Norm loss. The resultant predictor is a Dirichlet distribution on class probabilities which provides more detailed uncertainty model than the point estimate of the softmax-output deep nets (evidence predictor). The developed model results in increased accuracy and estimated uncertainty; however, linearly enhancing the size makes the deep ensemble unfeasible for memory-intensive tasks. To address the problem of memory intensive tasks, the model pruning and quantization techniques are performed along with deep ensemble and analyzed the effect in the context of uncertainty metrics. The magnitude-based weight pruning is used to prune the ResNet model, until it reached the required sparsity and a few epochs retrain the pruned networks. It is noticeable that increased disagreement in deep ensemble models will implies increased uncertainty which helps in making more robust predictions. The complete work is termed as Evidence-Efficient Compressed Deep Ensemble (EECDE) – ResNet which is appropriate for memory-intensive and uncertainty aware tasks. The test result show that the proposed model achieves accuracy of 98.29 % and 97.69% on CT and CXR mages for COVID-19 identification.