Artificial Intelligence-Assisted Cancer Status Detection in Radiology Reports

Area of Science:

• Medical Informatics
• Artificial Intelligence in Oncology
• Radiology Data Analytics

Background:

• Radiology reports contain valuable patient data for cancer research but lack structured formats.
• Manual curation of these reports is time-consuming, costly, and hinders research progress.
• Existing manual methods are sensitive to data volume, complexity, and require specialized skills.

Purpose of the Study:

• To develop and implement an AI-powered pipeline for efficient and accurate annotation of radiology reports.
• To leverage natural language processing (NLP) models for extracting critical data elements from unstructured text.
• To improve the speed, accuracy, and cost-effectiveness of data curation for cancer research.

Main Methods:

• Exploration of state-of-the-art AI techniques for automated data extraction.
• Implementation of an end-to-end pipeline for radiology report annotation.
• Training language models (LMs) as multiclass or multilabel classifiers to predict imaging sites, cancer presence, and status.

Main Results:

• Trained NLP models achieved high weighted F1 scores and accuracy in classification tasks.
• The AI-assisted curation process demonstrated increased accuracy and F1 scores compared to manual methods.
• Significant reduction in time and cost associated with data curation was observed.

Conclusions:

• AI and NLP models offer a faster and more accurate alternative to manual data extraction from radiology reports.
• Automated annotation assists manual curation, improving overall efficiency and reducing costs.
• This approach accelerates cancer research by enabling quicker access to structured patient data.