Radiology report generation for proximal femur fractures using deep classification and language generation models.

Area of Science:

• Artificial Intelligence in Radiology
• Medical Image Analysis
• Natural Language Processing for Clinical Reports

Background:

• Proximal femur fractures are a significant health issue in the elderly, impacting mortality, costs, and hospital stays.
• Accurate classification and diagnosis are crucial for effective patient management and resource allocation.
• Current methods often rely on manual data curation, limiting scalability and practical application.

Purpose of the Study:

• To develop and validate an automated method for subclassification of proximal femur fractures using X-ray images.
• To create an automated system for generating Dutch radiological reports for hip fractures.
• To eliminate the need for manually curated data in both classification and report generation models.

Main Methods:

• A fracture classification model was trained on 11,000 X-ray images from 5,000 electronic health records.
• A Dutch report generation model utilized embeddings from 20,000 pelvic fracture reports and was trained on 5,000 associated reports.
• The approach bypasses manual preprocessing of images and reports.

Main Results:

• The report generation model achieved performance on par with state-of-the-art methods using BLEU and ROUGE scores.
• User studies with medical students showed no significant difference between real and generated reports.
• Expert evaluation indicated generated reports approximate the quality of original reports.

Conclusions:

• The developed automated system for proximal femur fracture classification and Dutch report generation is highly promising for clinical practice.
• The method's independence from manual data curation significantly boosts its practical applicability.
• Challenges remain in generating sufficiently detailed and versatile training data for robust automated radiology report generation.