Natural Language Processing Techniques for Extracting Diagnostic Information from Radiology Reports

Main Article Content

B. Ranganatha Rao
K. L. Madhuri
D. Snehalatha

Abstract

The ability of extracting diagnostic information using Natural Language Processing (NLP) in unstructured radiology texts is becoming of increasing importance, and it is necessary due to the rapid growth of imaging studies. This paper dwells on automating the process of extracting diagnostic labels on English chest X-rays report texts by applying an NLP pipeline that comprises a hybrid of rule-based (chexpert, NegBio) and transformer-based (BioBERT, ClinicalBERT) components. This required the training and deployment of a machine-learning model process to parse radiology reports associated with a large-scale dataset, MIMIC-CXR-JPG and over 14 key pathologies and No Finding were attributed; with DICOM metadata also incorporated and structured to support downstream analysis with robustness. NLP systems showed high accuracy in prevalent findings like cardiomegaly, edema and pleural effusion (F1-scores >0.85), but displays reduced accuracy in the low frequency or linguistically confusing classes, and there is a severe class imbalance present: most findings are classified as No Finding. Exploratory data analysis and benchmarking indicated that rule-based tools are efficient when deployed on common patterns, whereas transformer models utilized to a larger extent contextual and uncertain information and required more substantial computational resources. The researchers summarize that hybrid NLP methods have the potential to automate and standardise the extraction of labels in reports of chest radiology, thus facilitating clinical records and enabling large-scale studies, as well as the improvement of the basis of AI-based diagnosis systems, although future research is required to solve the issue of generalisability and challenging linguistic constructions.

Article Details

Section
Articles