Deep-Learning Language-Modeling Approach for Automated, Personalized, and Iterative Radiology-Pathology Correlation

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Radiology-pathology correlation is crucial for medical education, quality assurance, and patient care.
• Traditional methods rely on manual review, which can be time-consuming and prone to error.

Purpose of the Study:

• To evaluate the efficacy of deep learning language models in matching radiology and pathology reports.
• To compare the performance of these modern AI techniques against previous correlation systems.

Main Methods:

• Utilized Universal Language Model Fine-Tuning for Text Classification (ULMFiT).
• Adapted models using 200,000 radiology and pathology reports.
• Trained a binary classification model on 100,000 report pairs (match/no-match).

Main Results:

• The deep learning model achieved 100% specificity, 65.1% sensitivity, and 73.7% accuracy.
• Performance closely mirrored a previous AI system (98.0% specificity, 65.1% sensitivity, 73.2% accuracy).
• The new model demonstrated rapid adaptation and comparable results to established methods.

Conclusions:

• Deep learning language models show significant promise for automated radiology-pathology correlation.
• These AI models offer continuous improvement and adaptability for clinical settings.
• The technology advances previous AI efforts, enabling performance tuning to user and site-specific needs.