Towards data-driven medical imaging using natural language processing in patients with suspected urolithiasis

Area of Science:

• Radiology
• Medical Informatics
• Natural Language Processing

Background:

• Most radiological reports are unstructured free text, hindering analysis.
• Manual evaluation of these reports is time-consuming and impractical for routine clinical use.
• Structured data is crucial for large-scale studies and quality assessment.

Purpose of the Study:

• To develop and apply a natural language processing (NLP) approach for automatically structuring narrative radiological reports.
• To extract clinical information and determine positive hit rates for suspected urolithiasis.
• To assess the correlation between clinical factors and confirmed urolithiasis.

Main Methods:

• Analysis of 1714 low-dose CT retroperitoneum reports using NLP.
• Automatic structuring of free-text reports based on RadLex concepts.
• Manual feedback for NLP engine training and validation; statistical analysis (chi-squared, logistic regression) to evaluate correlations.

Main Results:

• Urolithiasis was confirmed in 72% of reports.
• Kidney stones were noted in 38%, ureter stones in 45%.
• Previous stone history and obstructive uropathy strongly correlated with confirmed urolithiasis (p=0.001), with highest association found for previous stone history and loin pain (p<0.001).

Conclusions:

• NLP enables the conversion of free-text radiological reports into structured data.
• This structured data is valuable for epidemiological research and evaluating CT scan appropriateness.
• The approach facilitates answering diverse research questions using existing radiological report data.