CNN 아키텍쳐 기반 인공 고관절 판독 모델 성능 비교 연구

Abstract

Recently, in the process of total hip arthroplasty, due to the nature of artificial joints with different specifications from manufacturer to manufacturer, there is a chronic problem of using products compatible with products inserted in the past. To address this problem, the surgeon looks at the X-ray images before starting the operation and prepares the most similar implants based on empirical treatment. This preparation process may be subject to error depending on the physician's experience. As a result, many hip implant manufacturers around the world are actively developing image reading systems for implant products using their own data. Therefore, this study compares the performance of various CNN (Convolutional Neural Network) based architectural models that demonstrate superior performance in lesion diagnosis and prognosis prediction through high-dimensional abstraction based on visual information in medical image processing. Methods of study are as follows. X-ray images of the company's products provided by Correntec, the only artificial hip joint manufacturer in Korea, and the Cargle dataset 'Aseptic Loose Hip Implant X-ray Database' are used. In this case, the X-ray images used are not identified by the patient information. The acquired X-ray images are extracted from the implanted part, removed from the background, then subjected to data preprocessing such as binarization, removal of objects outside a particular region, and image inversion, and the preprocessed data set is converted to a binary classification model. The performance of each model is then compared by applying it to various architectural models based on CNN. There are a total of six CNN-based models used for performance comparisons: CNN, VGG-16, Google Net/Inception Net, Xception Net, Mobile Net, and ResNet, and the features and connection points of each model are described.