COVID-19 Classification by Integrating Lung Segmentation on CT Images

: This is a code repository of the one of the Finalists on 2020 INFORMS QSR Industry Data Challenge, CT Scan DIagnosis for COVID-19

Seonho Park (U. of Florida), Farnaz Babaie Sarijaloo (U. of Florida), Bijan Taslimi (U. of Florida)

Covid-19 pandemic is the most serious concern of this year, 2020. It is necessary to use effective and reliable methods to diagnose COVID-19. Molecular testing by nasal swab testing is one of the testing methods to disgnose COVID-19 but it still has a false negative issue. Computed tomography (CT) scans can be an auxiliary manner for screening and diagnosing COVID-19. This is a convolutional neural network based COVID-19 CT scan classification by integrating lung segmentation to augment the COVID-19 CT image data

Requirements

• pytorch == 1.5.0

Methodology

STEP 1: Lung segmentation

• ResUNet18 based lung segmentation
• Data: please download and place 2d_images.zip and 2d_masks.zip for the lung segmentation data from the link.
• Execution

python lungseg.py --datapath <datapath>

STEP 2: Classification

• MobileNetv2 is used as a backbone for the classification
• the input of the model is a grayscaled CT image as well as the lung segmentation output from the step 1
• Data: please download and place the following data from the link
• Execution

python train.py --datapath <datapath>

Results

Performance (AUROC & AUPR)

GradCAM++ Result