Keras is a neural network calculation library for Python environment. It enables implementation of deep learning related methods by relatively simple codes. Keras enables implementation of various methods from simple to state-of-the-art methods, such as methods proposed in the MICCAI conference. Also, it accelerates research and development on methods, which is important point for deep learning researches. Keras is quite useful library for research propose. This paper introduces how to implement an image classification program using convolutional neural network on Keras.

View full abstract

Caffe is one of deep learning frameworks, and has been developed with an easy optimization, high speed, and modularity. It is also open source and has active community. Therefore, it is easy framework for beginning the deep learning because users can be rich samples and reference models. For beginning the deep learning using Caffe, this paper describes how to install Caffe, how to implement an image classification, and how to implement the classification using transfer learning.

View full abstract

Chainer is an open source software library for deep learning. This paper shows the tutorial for deep learning using Chainer v2. The example of two-class figures (sphere or rectangular prism) classification task introduced in this paper is useful as a first step for resolving the difficult classification problems.

View full abstract

One of the most common tasks in medical imaging is semantic segmentation. Achieving this segmentation automatically has been an active area of research, but the task has been proven very challenging due to the large variation of anatomy across different patients. However, recent advances in deep learning have made it possible to significantly improve the performance of image recognition and semantic segmentation methods in the field of computer vision. Due to the data driven approaches of hierarchical feature learning in deep learning frameworks, these advances can be translated to medical images without much difficulty. Several variations of deep convolutional neural networks have been successfully applied to medical images. Especially fully convolutional architectures have been proven efficient for segmentation of 3D medical images. In this article, we describe how to build a 3D fully convolutional network (FCN) that can process 3D images in order to produce automatic semantic segmentations. The model is trained and evaluated on a clinical computed tomography (CT) dataset and shows stateof-the-art performance in multi-organ segmentation.

View full abstract

Deep learning can be applied to many tasks in medical image processing. There are some points to be careful in research use of deep learning to take advantage of its performance. Among such points, three topics including manual specification of many parameters in deep learning, pitfalls of data augmentation, and how randomly selected parameters in deep learning affect its performance, are discussed here.

View full abstract

Upon the progress of deep learning in various fields in recent years, use of deep learning is becoming common in medical imaging research. It is common to use GPU environment to speed up deep learning processing, but systematic information is not necessarily enough. In this paper, I will explain de basic information of GPU hardware and software environment. Then provide practical information of NVIDIA's useful solutions so that researchers can build optimal environment as easily as possible.

View full abstract

Medicine is moving toward “personalized medicine,” which is a novel concept for cancer treatment and prevention that takes into account individual variability (patient or tumor heterogeneity) in genes, lifestyle and environment for each person. However, there are several issues in the personalized medicine such as invasive biopsy, high cost and slow throughput for examination of gene mutations. Further, since tumors are heterogeneous, a small part of a tumor obtained by a single biopsy could not be reliable for the personalized medicine, and thus it could be difficult to carry out the personalized medicine in the cancer treatment. Therefore, radiomics concept has emerged for solving the issues and performing the “practical” personalized medicine. Radiomics is a novel field, which massively and comprehensively analyzes a large amount of medical images, and extracts mineable data that can make it possible to perform the personalized medicine. In this review paper, the authors describe what radiomics is, what radiomics can do, how to perform radiomics, advantages and disadvantages of radiomics, and the future of radiomics in cancer treatment.

View full abstract

Osteoporosis is known as a disease of bone. Visual screening using Computed Radiography (CR) images is an effective method for osteoporosis; however, there are many diseases that exhibit similar state of low bone mass. In this paper, we propose an automatic identification method of osteoporosis from phalanges CR images. As the proposed method, we implement a classifier based on Deep Convolutional Neural Network (DCNN) to identify unknown CR images as normal or abnormal. For training and evaluating of DCNN, we use pseudo color images. The pseudo color images are generated by assigning three types of ROI to R, G, and B channels after extracting the ROI from inside the phalange region of the three kinds of images created from the CR image. In the experiment, we apply our proposal method to 101 cases and True Positive Rate of 75.5 [%] and False Positive Rate of 13.9 [%] were obtained.

View full abstract

The author describes some foundation of construction of statistical shape models. A statistical shape model is constructed from a set of training medical images. The first operation needed for the model construction is to make correspondences between the training images. It is required to make correspondences between the surfaces of organs of different patients when one represents organ regions in images with Point Distribution model. If one employs levels sets for the region representation, then one needs to normalize the body shapes in the images in order to make correspondences between voxels in the images. The author introduces five categories of the methods for the making of the correspondences. In addition, the author briefly describes the relationships between statistical shape models and deep neural networks.

View full abstract