Abstract
In the medical field, various studies using artificial intelligence (AI) techniques have been attempted. Numerous attempts have been made to diagnose and classify diseases using image data. However, different forms of fracture exist, and inaccurate results have been confirmed depending on condition at the time of imaging, which is problematic. To overcome this limitation, we present an encoder-decoder structured neural network that utilizes radiology reports as ancillary information at training. This is a type of meta-learning method used to generate sufficiently adequate features for classification. The proposed model learns representation for classification from X-ray images and radiology reports simultaneously. When using a dataset of only 459 cases for algorithm training, the model achieved a favorable performance in a test dataset containing 227 cases (classification accuracy of 86.78% and classification F1 score of 0.867 for fracture or normal classification). This finding demonstrates the potential for deep learning to improve performance and accelerate application of AI in clinical practice.
Highlights
In the medical field, various studies using artificial intelligence (AI) techniques have been attempted
This retrospective study was approved by the Institutional Review Board (IRB) of Hanyang University Medical Center with a waiver of informed consent (HYUH 2019-06-003)
The dataset consisted of 786 X-ray images and 459 radiology reports and was split into training, validation (100 X-ray images), and test (227 X-ray images) sets, and the ratio of fracture to non-fracture cases was similar among datasets
Summary
Classification of bone fractures based on conventional machine learning pipelines consisting of preprocessing, feature extraction, and classification steps has been addressed. With the advent of deep learning models over recent years, several approaches to classify bone fractures have been proposed. Kazi et al.[14] attempted to classify proximal femurs based on AO classification standard, which is similar to the model proposed in the present study. These studies have demonstrated the potential of deep learning models, but they require large amounts of data. The model was validated using standard evaluation metrics including accuracy and F1 score, t-distributed stochastic neighbor embedding (t-SNE) to visualize representation vectors, and lesion visualization using gradient-weighted class activation mapping (Grad-CAM)
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
