Improved Automated Quality Control of Skeletal Wrist Radiographs Using Deep Multitask Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Radiographic Quality Control

Background:

• Radiographic quality control is crucial for accurate diagnosis and treatment planning in radiology.
• Manual quality assessment is time-consuming and prone to human error.
• Automated solutions are needed to improve efficiency and consistency in radiographic quality control.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated quality control of wrist radiographs.
• To classify key radiographic attributes: projection, laterality, presence of casts, and surgical hardware.
• To ensure congruence between image findings and electronic health record metadata.

Main Methods:

• Development of a multitask convolutional neural network (CNN) model using DenseNet 121 architecture.
• Training and validation on a dataset of 6283 wrist radiographs from 2591 patients.
• Evaluation of model performance using F1 scores for projection, laterality, cast, and hardware detection.

Main Results:

• High accuracy achieved in classifying projections (97.23%), detecting casts (97.70%), and identifying hardware (92.27%).
• Lower performance in laterality marker detection (82.52%), especially with partially visible markers.
• Demonstrated the potential of deep learning for automating key aspects of radiographic quality control.

Conclusions:

• The developed AI model shows significant promise for automating wrist radiograph quality control, improving workflow efficiency.
• Further refinement is needed to enhance laterality detection accuracy for comprehensive clinical application.
• Future research will focus on improving model robustness and expanding its utility in radiographic quality assurance.