In this paper, we introduce the background and the aim of our project supported by Japan Agency for Medical Research and Development (AMED). We also describe the structure and the operations of the cloud platform, the process management for the project, etc., and provides an overview of the subsequent papers.
Deep learning contributes to develop algorithms which obtain high accuracy on medical imaging. However, they focused on a certain disease and were not applied to various diseases because of difficulty in dataset construction. To address the problem, the AMED project was started collaborating with medical societies. In this paper, we report the progress and knowledge which we obtained through the project on endoscopic and pathological cancer detections.
We introduce our activities about biomedical image analysis. We also describe the purpose, method, and results about endoscopic image clustering that we have conducted in this project. In particular, we explain the refinement method of clustering results based on the repetition of supervision from a physician and image analysis. Furthermore, we introduce our ongoing projects regarding ulcerative colitis and duodenal papilla images.
This paper introduces our medical image analysis researches in the AMED project. Our research topics are radiological and endoscope image analyses. The radiological image analysis includes a blood vessel diagnosis assistance using non-contrasted abdominal CT volumes and an abdominal multi-organ segmentation from CT volumes. The endoscope image analysis includes observing area classifications of endoscope images taken in the stomach and colon. We explain approaches and results of these researches.
This article describes an image analysis project for extracting knowledge from a large scale heterogeneous (i.e., variations in facility, patient attributes, pathology and so on are very large) medical image database that several medical academic institutions supported by Japan Agency for Medical Research and Development (AMED) and National Institute of Informatics (NII) has been constructing. In this study, we specifically focus on the database organization and an example application in orthopedic surgery. The database organization uses an unsupervised learning framework by dimensionality reduction with t-SNE and k-means clustering. We selected the volumes that contains hip region from the heterogeneous large database and applied a deep-learning-based segmentation method to understand shape features of each anatomical part including bones and muscles. The statistical analysis of the shapes would allow us to understand musculoskeletal anatomy that helps diagnosis and planning in orthopedic surgery.
This study overviewed cardiac imaging techniques, applications, and future prospects. The heart actively deforms itself and its deformation drives the circulation of the body. In other words, there is a close association between morphology and function. Therefore, information on both morphology and function is helpful for diagnosing cardiac disease. In medical imaging research, many studies have examined how to obtain spatial, temporal, and functional information from the beating heart. As a result, comprehensive evaluation of the heart using multiple modalities has become common in clinical practice.
Osteoporosis is the main disease of bone. Although image diagnosis for osteoporosis is effective, there are concerns about increased burdens on doctors and variations in diagnostic results due to experience differences of doctors and undetected lesions. Therefore, in this paper, we propose a diagnostic support method to classify osteoporosis from Computed Radiography (CR) images of the phalanges and present classification results to doctors. In the proposed method, we constructed classifiers using Residual Network (ResNet), which is one type of convolution neural network, and classified the presence or absence of osteoporosis. For the input image to ResNet, we used the image generated from CR images. In this paper, we proposed three kinds of input images and conducted training and classification evaluation on each image. In the experiment, the proposed method was applied to 101 cases and evaluated using the Area Under the Curve (AUC) value on the Receiver Operating Characteristics (ROC) curve, the maximum value of which was 0.931.
Diffusion MRI is a powerful tool for characterizing the local properties of microstructures in living organisms with rich water molecules, especially for neuro brain area, using parameters of various signal models. A diffusion MRI dataset consists of signals measured using a variety of directions and strengths of the gradient field expressed in q-space. In addition to the well-known diffusion tensor imaging, a wide variety of signal models has been proposed for providing new information of local microstructures. In this series of short reviews for diffusion MRI, this manuscript covers introduction of diffusion weighted imaging, signal models and its parameter inference for better understanding of diffusion MRI basics.