Convolutional Neural Network-Based Automatic Analysis of Chest Radiographs for the Detection of COVID-19 Pneumonia: A Prioritizing Tool in the Emergency Department, Phase I Study and Preliminary "Real Life" Results

Abstract

The aim of our study is the development of an automatic tool for the prioritization of COVID-19 diagnostic workflow in the emergency department by analyzing chest X-rays (CXRs). The Convolutional Neural Network (CNN)-based method we propose has been tested retrospectively on a single-center set of 542 CXRs evaluated by experienced radiologists. The SARS-CoV-2 positive dataset (n = 234) consists of CXRs collected between March and April 2020, with the COVID-19 infection being confirmed by an RT-PCR test within 24 h. The SARS-CoV-2 negative dataset (n = 308) includes CXRs from 2019, therefore prior to the pandemic. For each image, the CNN computes COVID-19 risk indicators, identifying COVID-19 cases and prioritizing the urgent ones. After installing the software into the hospital RIS, a preliminary comparison between local daily COVID-19 cases and predicted risk indicators for 2918 CXRs in the same period was performed. Significant improvements were obtained for both prioritization and identification using the proposed method. Mean Average Precision (MAP) increased (p < 1.21 × 10−21 from 43.79% with random sorting to 71.75% with our method. CNN sensitivity was 78.23%, higher than radiologists’ 61.1%; specificity was 64.20%. In the real-life setting, this method had a correlation of 0.873. The proposed CNN-based system effectively prioritizes CXRs according to COVID-19 risk in an experimental setting; preliminary real-life results revealed high concordance with local pandemic incidence.

Keywords: Convolutional Neural Network (CNN); artificial intelligence; chest X-ray; coronavirus disease 2019 (COVID-19); deep learning; prioritization.

Figure 1

The tool architecture. The image (a) is pre-processed by algorithm (b) into the new image (c) which is elaborated by deep neural network (d) to extract features (e). Computer features are compared by the distance metric (i) with the previously extracted features (g) stored in the database (f) with their labels (h). The features (g) have been calculated on past known cases. Using the similarity information elaborated by (i) and labels (h) a COVID-19 score (m) is computed by stage (l). The set of most similar cases from the past (n), used for the computation of score (m), is returned to support doctor diagnosis and to provide interpretation of result.

Figure 2

The pre-processing stage. The original image (a) is properly resized limiting the maximum dimension to 1000 pixels but maintaining its original aspect ratio (b). The resized picture (c) is then processed by the U-Net model (d) to extract a binary mask (e) indicating where the lungs are located. To define the final cut (g), the binary mask (e) is framed by the smallest rectangle that contains the lungs (f). The resized picture (c) is normalized by mean and variance (h,i) and finally cropped (j) using the computed rectangular cut to obtain the final picture (k).

Figure 3

Data selection and labelling during phase 1. Images have been assigned to different groups according to RT-PCR test (for COVID-19) and assessment by a team of radiologists (for other diseases).

Figure 4

Data selection and elaboration process adopted in phase 2.

Figure 5

High level view on system installed at hospital for phase 2.

Figure 6

Daily new cases of COVID-19 in Provincia di Torino. Vertical axis: daily new cases, horizontal axis: date.

Figure 7

Distribution of the positive and negative groups.

Figure 8

Distribution of the predicted COVID-19 risk indicator for cases belonging to the three classes.

Figure 9

Distribution of the predicted COVID-19 risk indicator for cases belonging to the three classes, divided by gender. Blue boxes are referring to female, while orange ones are to male.

Figure 10

The 7 day rolling average comparison between AIppo COVID-19 score and Provincia di Torino daily new cases. Solid line indicates AIppo prediction, dots local new cases. Primary vertical axis (left): 7 day rolling average for AIppo score (in percentage), secondary vertical axis (right): 7 day rolling average for local new cases, horizontal axis: days.