Deep-COVID

Deep-COVID applies deep transfer learning to chest X-ray images to detect and classify COVID-19 infections.

Key Features:

• Dataset Preparation: Utilizes a curated dataset of 5,000 chest X-rays from publicly available sources annotated for COVID-19 presence by board-certified radiologists, with a 2,000-image training subset and a 3,000-image evaluation set.
• Deep Learning Models: Employs transfer learning to fine-tune four convolutional neural networks—ResNet18, ResNet50, SqueezeNet, and DenseNet-121—on the training subset.
• Performance Metrics: Reported model performance on the evaluation set includes sensitivity ≈98% (±3%) and specificity ≈90%.
• Evaluation Techniques: Provides receiver operating characteristic (ROC) curves, precision-recall curves, average prediction scores, and confusion matrices for each model.
• Heatmap Generation: Produces heatmaps that highlight lung regions potentially infected by COVID-19 and that align with radiologist annotations.
• Implementation Framework: Model implementations are provided in PyTorch.

Scientific Applications:

• Rapid Screening and Diagnosis: Supports rapid screening and diagnosis of COVID-19 from chest X-rays to assist clinical prioritization and isolation decisions when RT-PCR testing is limited or delayed.

Methodology:

Deep transfer learning was applied by fine-tuning pre-trained ResNet18, ResNet50, SqueezeNet, and DenseNet-121 models on a 2,000-image training subset and evaluating on the remaining 3,000 images, with assessment via sensitivity, specificity, ROC and precision-recall analyses and generation of model heatmaps.