A scoping review of large language model based approaches for information extraction from radiology reports

Abstract

Radiological imaging is a globally prevalent diagnostic method, yet the free text contained in radiology reports is not frequently used for secondary purposes. Natural Language Processing can provide structured data retrieved from these reports. This paper provides a summary of the current state of research on Large Language Model (LLM) based approaches for information extraction (IE) from radiology reports. We conduct a scoping review that follows the PRISMA-ScR guideline. Queries of five databases were conducted on August 1st 2023. Among the 34 studies that met inclusion criteria, only pre-transformer and encoder-based models are described. External validation shows a general performance decrease, although LLMs might improve generalizability of IE approaches. Reports related to CT and MRI examinations, as well as thoracic reports, prevail. Most common challenges reported are missing validation on external data and augmentation of the described methods. Different reporting granularities affect the comparability and transparency of approaches.

Fig. 1. PRISMA flowchart describing the source of evidence retrieval and selection process.

Querying of five databases resulted in a total of 1237 sources of evidence eligible for screening. This number was reduced to 374 after deduplication and removal based on publication year. Eventually, 34 studies were included in this review after completion of the screening process.

Fig. 2. Distribution of reported NLP tasks.

The circles contain the absolute number of studies per task. NER Named entity recognition, RE Relation extraction, ML-CL Binary multi-label classification, MC-CL Multi-class classification, QA Question answering.

Fig. 3. Distribution of modalities.

The diagram shows absolute numbers of mentioned modalities. Several studies use reports obtained from multiple modalities. Other modalities include positron emission tomography-computed tomography (PET-CT) (n = 1) and ultrasound (n = 2). Three studies did not explicitly mention associated modalities. Abbreviations: CT Computer tomography, MRI Magnetic resonance imaging.

Fig. 4. Distribution of anatomical regions.

The diagram shows absolute numbers of mentioned anatomical regions. Several studies use reports corresponding to multiple anatomical regions. Other anatomical regions include the heart, abdomen, pelvis, “all body regions'', nose, thyroid (n = 1 each) and breast (n = 2). Four studies did not explicitly mention associated anatomical regions.