Purpose/Objectives
We present a deep learning model built using transfer learning, which can detect COVID-19 pneumonia patterns from digital chest radiographs, perform “triage” and flag these as priority scans needing immediate attention. It can recognize patterns and undertake complex computational tasks often far rapidly and with increased precision than a human reader. Using artificial intelligence in imaging practice can also ensure study list re-prioritization ensuring quick attention to urgent studies.
Methods & Materials
We used multiple open and private hospital-based Institutional test datasets with a separate “hold-out” test set to ensure both internal and external validity. A fixed part of the data (20%) was upsampled with various augmentation methods to induce data variability and ensure model robustness. The test comprises of 1050 COVID positive patients and 1245 patients with no chest abnormality. The deep learning models were based on U-Net architecture with Spatial attention Gate and Attention encoder. A pixel-wise probability threshold was used to convert the segmentation output into a binary classification label.
Results
Out of multiple models validated finally, three were tested on the external hold-out test dataset. To reduce bias in performance assessment, as there is no true gold standard for COVID detection, we compared the performance of the models with the three radiologists separately by adjusting the model’s sensitivity as per individual radiologist. The models gave a performance of 0.83, 0.86 and 0.87 area under the curve, respectively, with model-3 (DxCOVID) performing best. There was no significant statistical difference between the performance of model 3 when compared to two of the three readers.
Conclusion
While the practice of reading pneumonia from chest radiographs is age-old with many experienced radiologists doing this regularly, it is not the case for COVID-19. COVID-19 radiographs have particularly proved to be relatively difficult to read, particularly for those with a mild form of the disease. This is a primary cause of the overall low sensitivity of chest radiography reporting. In the current situation where the healthcare providers are overburdened because of the caseload and are also dealing with a lack of resources needed for large-scale image interpretation, AI could serve as a great assistant in the reading process. The findings have implications for screening and triage, initial diagnosis, monitoring disease progression, Thus, the utilization of AI can help support institutes and organizations which are swamped due to heavy caseloads and suffer from an overall lack of sufficiently experienced radiologists.
Disclosure statement
Dr Amit Kharat, one of the co-authors is the Professor at the Institution where the study was performed and also a co-founder of the AI firm whose expertise was needed for building the model.
Abstract Type
Oral Presentation
Presenter's firstname and lastname
Pranav Ajmera
Presenter's city and country
Pune, India
Presenter's email address
pranavajmera@gmail.com

PRE-PUBLISHING: I confirm that this work has not been published or presented at any other meeting prior to December 2021.
I confirm.
PUBLICATION: I accept that ESTI may publish my abstract in an online syllabus and may display the abstract at the ESTI website during and/or after the ESTI Winter Course 2021.
I accept.
CONGRESS REGISTRATION: I accept that the presenter of this abstract has to register for the ESTI Winter Course 2021, in order to present the abstract and/or to have the digital poster displayed during and/or after the course in the online system.
I accept.
AUTHORS AGREEMENT: I confirm that all co-authors mentioned in the author block of this abstract have been informed about, and agreed to this submission.
I confirm.
MATERIAL USED: I confirm that this abstract does not contain any material that is libellous, defamatory, or otherwise unlawful. It does not contain any material that invades the right of privacy, any proprietary or copy-rights.
I confirm.

A Single Centre Institutional Study on 2295 patients for COVID-19 Detection from Radiographs Using a Deep Learning Model - Our Experience and Insights

The Winter Course 2021 covers a variety of topics from basics in interstitial lung disease, infections including Covid-19, malignancies, lung cancer screening, and artificial intelligence. The combination of topics is aimed to serve as a good introduction to radiologists embarking on a career in thoracic imaging as well as a solid refresher for experienced thoracic radiologists.

[ESTI Winter Course 2021 Evaluation Form](https://www.surveymonkey.com/r/PP9WFTJ?sid=%5Bsid_value%5D&uid=%5Buid_value%5D&tkn=%5Btkn_value%5D&cgrid=%5Bcgrid_value)

ESTI Winter Course 2021

A Computed Tomography-based Radiomics Model for Patient Prognostic Stratification in Non-Small Cell Lung Cancer

Break Video

CT-Guided Microcoil placement for VATS Resection of Pulmonary Nodules

Development of a risk prediction model for pneumothorax in percutaneous CT guided lung biopsy

Evaluation of an AI program in thoracic imaging: Performance in detection of thoracic aorta ectasia

Evaluation of AI applications in thoracic imaging: Performance in the detection of vertebral body alterations

Can inclusion of the combination of chemical shift ratio and signal intensity increase the number of patient’s classified as definitively benign on thymic MRI beyond use of chemical shift ratio alone? Experience from a UK thoracic tertiary referral centre

Relative contribution of emphysema and airway disease to disease severity in COPD patients GOLD stages I to IV using quantitative computed tomography

Lung Thermal Ablation: Comparison between an Augmented Reality Computed Tomography (CT) 3D Navigation System (SIRIO) and Standard CT-Guided Technique

Differentiating COVID-19 pneumonia from atypical bacterial, fungal and non-COVID-19 viral pneumonia on Chest-CT: Is it possible?

Reliability of Ultra-Short Echo Time MRI sequence in the assessment of Interstitial Lung Disease related to Systemic Sclerosis

Assessing longitudinal changes in lung gas transfer in dyspnoeic post-COVID-19 patients using hyperpolarized Xenon-129 Magnetic Resonance Imaging

Session Introduction