The Application of Deep Learning Methods for Proximal Humerus Fracture Feature Identification

点击翻译按钮获取中文摘要

SUMMARY Proximal humerus fractures (PHFs) are the third most common fracture in the elderly, with an estimated 200,000 occurring each year in the United States. PHFs can lead to pain, poor shoulder function, plus short and long- term disability for patients. Substantial controversy persists regarding initial treatment for elderly adults with this injury. PHFs can be managed conservatively or surgically and great controversy exists over which patients should be treated surgically. A unique challenge with PHFs is that they have variable presentation and range in complexity. Unlike the management of other major joint fractures, the initial treatment choice for PHF is highly dependent on the fracture characteristics. Treatment effectiveness evidence is needed to guide clinical care for individual patients with PHF. The Neer Classification, first developed in 1970, is the most widely used framework to describe and classify PHFs. Although the Neer Classification is the most widely used in practice, it is outdated, incomplete, often incorrectly applied, and suffers from poor interobserver reliability. The absence of a universally accepted, standardized fracture classification system is a critical barrier in the development of treatment effectiveness evidence for PHF. The application of deep learning (DL) computational models can automate and standardize the fracture classification process and identify all relevant fracture characteristics. DL image analysis models have been shown to be highly accurate at identifying features of interest on diagnostic images. An automated, standardized PHF classification system will enhance our ability to universally standardize fracture classification across all orthopaedic clinical care settings, improve the precision and efficiency in fracture care and generate treatment effectiveness evidence to guide clinical practice. The overall objective for this application, is to develop and validate a DL computational model capable of identifying fracture features using X-ray images. Our central hypothesis is that we can develop a DL model that will be as accurate as expert shoulder specialists in identifying important fracture features on X-ray images. In Aim 1 we will modify the Neer Classification framework for fracture feature identification. Aim 2 will be the development of a gold standard dataset for deep learning DL fracture feature identification, and finally Aim 3 will be the training and testing of a DL model to identify fracture features on X-rays.