| 소속 |
Dankook University Hospital, Department of Rehabilitation Medicine1, Dankook University, Department of Nanobiomedical Science & BK21 PLUS NBM Research Center for Regenerative Medicine2, Dankook University, Institute of Tissue Regeneration Engineering (ITREN)3, Dankook University, Department of Applied Computer Engineering4, Dankook University, Department of Software Engineering5 |
Objective
Over the past few decades, medical imaging techniques have been used for the early detection, diagnosis, and treatment of diseases. Conventionally, those images are interpreted by human experts such as radiologists and physicians. However, given wide variations in pathology and the potential fatigue of human experts, researchers and doctors have begun to benefit from computer-assisted interventions. Videofluoroscopic swallowing study (VFSS), which is one of the radiologic studies that demands high concentration and fatigue of reading physicians, could be a good candidate. We aimed to make a deep learning program that can automatically detect cervical spines from VFSS images, as a first step in development of VFSS reading system.
Methods
From 195 VFSS video files, images were separated in 24 frame per second. Images most distinct cervical spine in each file were selected and calibrated via contrast limited adaptive histogram equalization (CLAHE) technique. After creating a ground truth for each image based on manual annotation of vertebrae boundaries, image patches of 120 x 120 pixel size centered on ground truth were extracted following Gaussian distribution. Image patches overlapping with ground truth by 45% or more were defined as positive samples, and those overlapping less than 25% were treated as negative samples. Samples overlapping 25-45% were excluded because they were likely to act noise. Through this procedure, about total 69,400 samples were obtained. Machine learning was performed by three models of convolution neural network (one conventional machine learning model (support vector machine with histogram of oriented gradient (Hog + SVM)) and two deep learning models (Alexnet, and Resnet 50)) and the accuracy was compared.
Results
Cervical detection rate of each model delineated as following order (SVM, Alexnet, Resnet 50). Accuracy was 92%, 96%, and 99%, recall was 94%, 95%, and 99%, Precision was 90%, 98%, and 99%, respectively.
Conclusion
By deep learning, computer can detect cervical spine accurately from videofluorospic images without any manual intervention. Resnet 50 showed best result among the models used in this study. Authors believe that the results of this study will contribute to development of automatic reading program for VFSS.
|
