CXR-Net: An Artificial Intelligence Pipeline for Quick Covid-19 Screening of Chest X-Rays
Haikal Abdulah, Benjamin Huber, Sinan Lal, Hassan Abdallah, Luigi L. Palese, Hamid Soltanian-Zadeh, Domenico L. Gatti
CXR-Net is a two-module Artificial Intelligence pipeline for the quick detection of SARS-CoV-2 from Antero/Posterior (A/P) chest X-rays (CXRs). Module 1 was trained on a public dataset of 6395 A/P CXRs with radiologist annotated lung contours to generate masks of the lungs that overlap the heart and large vasa. Module II is a hybrid convnet in which the first convolutional layer with learned coefficients is replaced by a layer with fixed coefficients provided by the Wavelet Scattering Transform (WST). Module 2 takes as inputs the patients’ CXRs and corresponding lung masks calculated by Module I, and produces as outputs a class assignment (Covid vs. non-Covid) and high resolution heat maps that identify the SARS associated lung regions. Module 2 was trained on a dataset of CXRs from non-Covid and RT-PCR confirmed Covid patients acquired at the Henry Ford Health System (HFHS) Hospital in Detroit. All non-Covid CXRs were from pre-Covid era (2018-2019), and included images from both normal lungs and lungs affected by non-Covid pathologies. Training and test sets consisted of 2265 CXRs (1417 Covid –, 848 Covid +), and 1532 CXRs (945 Covid –, 587 Covid +), respectively. Six distinct cross-validation models with the same Covid +/– ratio, each trained on 1887 images and validated against 378 images, were combined into an ensemble model that was used to classify the CXR images of the test set with resulting Accuracy = 0.789, Precision = 0.739, Recall = 0.693, F1 score = 0.715, ROC(auc) = 0.852.
Source code for CXR-Net is deposited at: https://github.com/dgattiwsu/CXR-Net
The repository contains two folders, Module_1 and Module_2, with all the jupyter notebooks that were used for training and validation. There are only example images for inputs and outputs. The folder run_CXR-Net contains two jupyter notebooks, Predict_new_patient_from_processed_png.ipynb and Predict_new_patient_from_unprocessed_dcm.ipynb, plus additional folders that can be used to run the entire CXR-Net pipeline with the provided example image. Please notice that with the release of Kymatio (0.3.dev0) used to prepare this version of CXR-Net, the ensemble model must be generated by first compiling separately the Wavelet Scattering Transform model and then joining it to the ensemble resnet model. The next release of Kymatio will allow the entire ensemble model to be saved on file and reloaded in memory when needed. Currently, the ensemble model compilation takes about 2 minutes, but once the model is loaded in memory, the class prediction (Covid positive or negative, with corresponding scores) and the heat map generation take about 2 seconds per image with 2 V100 GPUs.
A Python script, Predict_new_patient.py, can be used to select from the command line both source/target directories, and, optionally, which dcm image files must be processed for covid/non-covid classification.
Refer to the ArXiv manuscript for additional information on how the JMS and V7 databases of CXRs with segmented lungs were generated. These databases are available on request from the corresponding author.
The database of Covid negative and Covid positive CXRs used in this study is an exclusive property of the Henry Ford Health System in Detroit, and is not publicly available.
g-CXR-Net is a graphic interface to run CXR-Net on Mac, Linux, or Windows desktops. A brief medRxiv article describing the features of g-CXR-Net is available at https://medrxiv.org/cgi/content/short/2021.06.06.21258428v1, or by mobile device's via the QR code shown on the side. Source code for g-CXR-Net is deposited at: https://github.com/dgattiwsu/g-CXR-Net
This graphic application can be launched from a virtual env with the command line syntax: python g_CXR_Net.py --xdir 'executable_dir', where 'executable_dir' is the directory where g_CXR_Net.py itself is located. This syntax is retained in order to facilitate the creation of an icon located anywhere in the os tree, linked to the application. Python scripts to generate icon links to run g-CXR-Net are provided in the directory CXR_Net_icons. The app can also be launched from outside a virtual environment, but in this case a full path to python location in the virtuel env must be provided.
A short video describing the installation and use of g-CXR-Net is available HERE.
