GraphCovidNet

GraphCovidNet applies Graph Isomorphic Network (GIN) graph neural network methods to CT scans and chest X-rays (CXRs) by converting images into undirected graphs to detect COVID-19.

Key Features:

• Graph Isomorphic Network (GIN): Implements a GIN-based architecture for learning from graph-structured representations of medical images.
• Imaging modalities: Operates on CT scans and chest X-rays (CXRs) as input image types.
• Graph conversion preprocessing: Converts image data into undirected graphs during preprocessing to represent structural information.
• Edge-focused representation: Emphasizes graph edges rather than entire raw images to capture relevant structural features for classification.
• Datasets evaluated: Evaluated on the SARS-COV-2 Ct-Scan dataset, COVID-CT dataset, a combination of the covid-chestxray-dataset and Chest X-Ray Images (Pneumonia) dataset, and the CMSC-678-ML-Project dataset.
• Performance: Reported overall accuracy of 99% across these datasets and perfect accuracy in binary classification tasks distinguishing COVID-19 from non-COVID-19 scans.

Scientific Applications:

• COVID-19 detection in medical imaging: Classifies CT and CXR scans to identify COVID-19-related findings.
• Binary classification tasks: Separates COVID-19 scans from non-COVID-19 scans with reported perfect binary accuracy.

Methodology:

Images are preprocessed into undirected graphs that emphasize edges, and a Graph Isomorphic Network (GIN) is trained on those graph representations; the model was evaluated on the SARS-COV-2 Ct-Scan, COVID-CT, covid-chestxray combined with Chest X-Ray Images (Pneumonia), and CMSC-678-ML-Project datasets.