Improved Diagnosis of COVID-19 from Chest X-Rays Using Local Phase-Based Image Enhancement and Deep Learning

摘要

The COVID-19 pandemic continues to have a devastating effect on the health and well-being of the global population. Under the globe COVID-19 crisis, public health care systems have confronted challenges in many aspects including a critical shortage of medical resources. In fighting against COVID-19, effective diagnosis and triaging of infected patients is critical for preventing the spread of diseases and providing adequate care. Radiological imaging, such as Computed Tomography or Chest X-ray (CXR), has been used extensively. CXR due to its faster imaging time, wide availability, low cost, and portability gained much attention. To reduce intra- and inter-observer variability, during the radiological assessment, and improve diagnostic time computer-aided computational tools have been developed to supplement medical decision-making and subsequent management. Supervised deep learning, which is a popular research area of artificial intelligence, enables the creation of end-to-end models to achieve promised results and to provide timely assistance to patients. However, the performance of such models relies on the availability of a large and representative labeled dataset. The creation of which is a heavily expensive and time-consuming task, and especially imposes a great challenge for a novel disease, like COVID-19. Semi-supervised learning and self-supervised learning have shown the ability to match the incredible performance of supervised models whilst requiring a small fraction of the labelled examples. This makes the semi or self-supervised paradigm an attractive option for identifying COVID-19.The goal of this thesis work is to develop robust, accurate, and automatic CXR classification method for COVID-19 diagnosis. First, an enhanced CXR representation is generated using a local phase-based image enhancement approach. A novel multi-feature Convolutional Neural Network (CNN) architecture, which is guided by both original CXR and enhanced CXR, is developed for improving COVID-19 diagnosis. Next, a Parallel-Attention block based on the self-attention mechanism is developed and applied to the proposed multi-feature CNN for fusing the features at different spatial resolutions.To solve the issue of limited labeled data and to provide an alternative for the tedious labeling process, we introduce a multi-feature-based deep semi-supervised pipeline for classifying COVID-19from CXR. Our pipeline is based on a teacher/student paradigm, that leverages a large number of unlabeled images. We demonstrate, through our experiments, that our model is able to outperform the current State-of-The-Art supervised model with a small fraction of the labeled examples. In the end, we propose a self-supervised learning method, termed MoCo-COVID, which is an adaption of the contrastive learning method MoCo, to produce models with better representations and initializations for the detection of COVID-19 in CXR. We find that MoCo-COVID pretraining provides the most benefit with limited labeled training data. We then propose a new Vision Transformer-based multi-feature architecture using cross-attention mechanism for COVID-19 diagnosis and show that this model achieves an improved accuracy with a small fraction of labeled data. We evaluate our methods on the largest COVID-19 dataset.