SARS-COV-2 constitutes a novel virus with a relatively high infection rate [2]. As such, early and effective detection is needed in order to track and impede the spread of the pandemic. We propose a Deep Learning approach for detecting COVID-19 based on X-ray images. Aside from detection, our network provides insight about the affected area, the core locations of infection and how similar diseases are when represented by networks. Our findings illustrate promising results both in localisation as well as AUC, reaching a ROC curve performance of 0.88 on the COVID-19 X-ray Dataset. Our report also analyses the importance of Transfer Learning, Model Architecture and Data Pre-processing.
Our model builds upon well established architectures for robustness and efficacy. It consists of a core deep Convolutional Neural Network which is augmented with a fully connected layer forclassification. Mainly two types of architetures were explored: DenseNet121 and ResNet50.
Our Deep Neural Network approach produced noteworthy results, showcasing great potential androbustness in diagnostic tasks. Its latent representation was able to properly separate the classesand created distinct clusters, by producing a new vector space that can model sample characteristicsmore accurately. Furthermore, we showed that the network can be utilized for other tasks, not justdiscriminative, by applying SOMs to uncover class relationships and using heatmaps (CAM) tohighlight the affected region. Using the latter, the network’s decision can be investigated to avoidoverfitting due to data set particularities (e.g. small and unbalanced data sets). Our experiments showthat an effective way of avoiding this is using pre-trained models, as they initialize the networks inbetter regions of the learning space, avoiding such undesirable overfit solutions.
Detailed report availiable here
To run the code the main.py should be executed. To select the database import the appropriate dataset object. Keep in mind that the path should be adjusted in the object file and for the chestXray14 dataset, the images should be downloaded!