WriVision: A Machine Learning Model for Quality Control of Wrist X-Rays

Overview

Radiologists are heavily reliant on X-rays for precise medical diagnosis. However, the acquisition of suboptimal views during X-ray imaging can lead to misdiagnosis, potentially putting patient care at risk. T o address this issue, we designed and trained an advanced artificial intelligence model that can provide quality control for technologists during X-ray acquisition. 55 Our approach involved creating a Convolutional Neural Network based on the DenseNet121 architecture that could identify the true projection, laterality, and presence of cast or hardware on our dataset of 6823 de-identified patient radiographs. Our model was trained in batches of 64 images with 30 epochs, utilizing a learning rate of 1x10-3, a decay of 0.3 with a minimum threshold of 1x10-6 and a dropout of 0.1 to prevent overfitting. T o test our model, we utilized a 5-fold stratification validation, allowing us to evaluate the model’s efficacy on every image in our dataset. Our results demonstrated an impressive F1 score of 93.1 for projection, 78.7 for laterality, 96.83 for the presence of cast, and 92.12 for the presence of hardware. Our team is currently working on developing a software that integrates our model into existing X-ray software solutions such as dicomPACS. This software will provide technologists with real-time predictions for the four classes, alerting them if any issues arise and allowing for prompt corrective action, including rescanning the patient. By ensuring optimal views during X-ray acquisition, our innovative model has the potential to greatly enhance patient care.