Structured report data can be used to develop deep learning algorithms: a proof of concept in ankle radiographs

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning models require large, accurate datasets for training.
• Current methods for label extraction involve time-consuming manual review or natural language processing.
• There is a need for efficient methods to generate training data for medical AI.

Purpose of the Study:

• To develop and evaluate a workflow for using structured radiology reports as labels for deep learning.
• To automate the process of data acquisition and preparation for AI model training.

Main Methods:

• Included anterior-posterior ankle radiographs with available structured reports.
• Developed a script to automatically retrieve, convert, and anonymize radiographs from PACS.
• Retrained a deep convolutional neural network using data from structured reports.
• Evaluated model performance on an independent set of radiographs.

Main Results:

• The automated workflow, once configured, completed in under one hour.
• Successfully retrieved 157 structured reports and corresponding radiographs.
• The deep learning model achieved an area under the curve of 0.850 for fracture detection on an unseen dataset.

Conclusions:

• Data from routine clinical structured reports can successfully train deep learning algorithms.
• Structured reporting holds significant potential for advancing AI in radiology.
• This approach streamlines the development of AI tools for medical image analysis.