[Paper Review] Identifying Radiological Findings Related to COVID-19 from Medical Literature

Yuxiao Liang, Pengtao Xie|arXiv (Cornell University)|Apr 4, 2020

COVID-19 diagnosis using AI9 references6 citations

TL;DR

This paper proposes a BERT-based natural language processing framework to extract and reconcile radiological findings related to COVID-19 from a global corpus of medical literature. By applying sentence classification and constituent parsing to the CORD-19 dataset, the method identifies high-frequency, clinically relevant abnormalities such as ground-glass opacities, pleural effusion, and interlobular septal thickening, offering a globally representative, unbiased summary of COVID-19 radiological features.

ABSTRACT

Coronavirus disease 2019 (COVID-19) has infected more than one million individuals all over the world and caused more than 55,000 deaths, as of April 3 in 2020. Radiological findings are important sources of information in guiding the diagnosis and treatment of COVID-19. However, the existing studies on how radiological findings are correlated with COVID-19 are conducted separately by different hospitals, which may be inconsistent or even conflicting due to population bias. To address this problem, we develop natural language processing methods to analyze a large collection of COVID-19 literature containing study reports from hospitals all over the world, reconcile these results, and draw unbiased and universally-sensible conclusions about the correlation between radiological findings and COVID-19. We apply our method to the CORD-19 dataset and successfully extract a set of radiological findings that are closely tied to COVID-19.

Motivation & Objective

• To address inconsistencies in radiological findings for COVID-19 due to population bias in hospital-specific studies.
• To develop an automated, scalable method for identifying and reconciling radiological findings across diverse global medical literature.
• To reduce reliance on manual annotation by using a bootstrapping approach for training a sentence classifier with minimal human labeling.
• To produce a comprehensive, frequency-ranked list of radiological findings that are most strongly correlated with COVID-19 for clinical evaluation.

Proposed method

• A BERT-based sentence classifier is trained to identify sentences in medical literature that describe radiological findings related to COVID-19.
• A bootstrapping strategy is employed to iteratively identify and label hard negative examples, improving classifier performance with minimal manual annotation.
• Constituent parsing is applied to positive sentences to extract noun phrases representing radiological abnormalities, such as 'ground-glass opacity' or 'pleural effusion'.
• Frequency analysis is performed on extracted noun phrases to rank the most prevalent findings across the dataset.
• The method aggregates findings from over 2,000 articles in the CORD-19 dataset, spanning diverse geographic and demographic populations to reduce bias.
• The final output provides a ranked list of high-frequency, clinically relevant radiological findings for further medical validation.

Experimental results

Research questions

• RQ1 Which radiological findings in lung imaging are most consistently reported across global medical literature as associated with COVID-19?
• RQ2 How can natural language processing be used to reconcile conflicting or inconsistent findings from isolated hospital-based studies?
• RQ3 To what extent can a BERT-based classifier with bootstrapping achieve high precision in identifying radiological findings from unstructured medical text with minimal human-labeled data?
• RQ4 What are the most frequently reported radiological abnormalities in COVID-19 patients when aggregated across diverse populations and institutions?

Key findings

• The method successfully extracted 15 high-frequency radiological findings from over 360,000 sentences in the CORD-19 dataset, with ground-glass opacities and pleural effusion among the most prevalent.
• The BERT-based classifier achieved high precision in identifying sentences containing radiological findings, with performance enhanced through bootstrapping on hard examples.
• Consolidation, ground-glass opacities, interlobular septal thickening, and pleural effusion were consistently identified as key imaging features across multiple studies.
• The ranked list of findings, derived from a globally diverse dataset, provides a more objective and universally representative summary than isolated hospital studies.
• Example sentences from the dataset confirmed that findings like 'crazy-paving pattern' and 'subpleural cystic lesions' are frequently associated with COVID-19 in clinical reports.
• The approach validated known clinical associations, such as fibrosis appearing in follow-up scans and bronchiectasis emerging in later disease stages, aligning with literature from Kanne et al. (2020) and others.