ISAKOS: 2023 Congress in Boston, MA USA

2023 ISAKOS Biennial Congress In-Person Poster

 

Automated Detection of Traumatic Hand Fractures on Plain Radiographs Using a Deep Learning Model

Logan Nye, MD, Allston, MA UNITED STATES
Soheil Ashkani-Esfahani, MD, Boston, Massachusetts UNITED STATES
Hamid Ghaednia, PhD, Boston, MA UNITED STATES
Joseph Hasbrouck Schwab, MD, Boston, Ma UNITED STATES

Massachusetts General Hospital, Boston, MA, UNITED STATES

FDA Status Not Applicable

Summary

A machine learning model trained on trauma radiographs demonstrates ability to localize acute hand fractures on plain film.

Abstract

Objectives:
Acute hand traumas are common orthopaedic injuries in athletic competition. Healthcare providers that suspect their athlete sustained a fracture use plain radiographs to closely survey bones in the hand. Despite their common nature, some acute hand injuries - particularly occult fractures of the carpals and phalanges - are difficult to visualize on plain films and may go undiagnosed. Deep learning modalities have demonstrated potential for distinguishing subtle fracture patterns in medical imaging. These technologies may help prevent missed fracture diagnoses in ambulatory settings. Here, we investigate this use of artificial intelligence models for acute hand trauma. Our objective was to train and evaluate a deep learning algorithm capable of detecting and localizing hand fractures on plain films.

Methods

Our team compiled a dataset of 1548 hand trauma radiographs, annotating all fractures with a bounding box. The images were uniformly resized to 412x412 pixels and randomly divided into training (70%), validation (20%), and test (10%) data subsets. The training subset was augmented with affine image transformations to create 3 training outputs per original image, resulting in an image dataset of 3.2k training samples, 314 validation samples, and 154 test samples. A neural network was trained on this dataset for 550 epochs. Metrics including average precision (mAP), loss function, overall precision, and recall were evaluated during the training period.

Results

Overall average precision on validation data and test data were 88% and 89%, respectively. The trained model had a mAP of 88.0%, precision of 91.0%, and recall of 86.1%. The dataset and a hosted web API of the model are available for download and use (https://universe.roboflow.com/logans-space/hands-u38uw/model/1).

Conclusions

This preliminary study suggests deep learning modalities may be used to develop fracture detection adjuncts. Algorithms such as this demonstrate potential for improved fracture detection in high acuity healthcare environments like the emergency department.

Keywords: Artificial Intelligence, Imaging, Machine Learning, Orthopaedics, Sports Medicine