S.D. Dias, P.E.M. Lopes
FastCompChem, Lda,
Portugal
Keywords: COVID-19, deep learning, diagnosis, chest X-ray, machine learning
Summary:
Accurate diagnosis of COVID-19 infection is a crucial tool to identify and control the spread of the disease. Current frontline technology, based on Reverse Transcription Polymerase Chain Reaction (RT-PCR), can depend on getting an adequate sample from the patient. If the sample collected does not have enough viral material on it, either because of the collection method or because of the stage of the infection, it can affect the result. RT-PCR also requires well equipped laboratories and specialized human resources. The technology we developed uses deep learning to analyse chest x-Rays with the goal of identifying signs of disease at a very early stage. Proprietary deep learning algorithms, designated FastNet, that are based on TensorFlow (www.tensorflow.org) were developed using a toy dataset [1]. The results from the toy dataset are excellent (sensitivity: 0.984; accuracy: 0.998). Training involved radiographs of COVID-19 and pneumonia infected patients and of healthy individuals. These results, which are arguably too good, will be confirmed in a retrospective study, to be started soon, in the Portuguese National Health System (SNS). The clinical trial will compare FastNET analysis with RT-PCR confirmed cases. In a second study, the performance of FastNET will be assessed relative to the diagnosis of expert radiologists for specific application in emergency rooms. Recently, Tizhoosh and Fratesi raised questions about the use of toy datasets [2]. We totally agree with their view, but also find the use of toy datasets of great value in early development and testing. The technology has multiple uses, but two specific uses are being pursued immediately: (1) in emergency rooms for fast triage of patients with symptoms compatible with COVID-19 infection and (2) to the armed forces in specific settings, for example Navy ships or deployments to remote locations. A specific web interface (see Figure 1) and a standalone Windows program have been developed. [1] Cohen JP, Morrison P, Dao L (2020) COVID-19 image data collection. arXiv 200311597 [2] Tizhoosh HR, Fratesi J (2021) COVID-19, AI enthusiasts, and toy datasets: radiology without radiologists. Eur Radiol 31:3553-3554