To date, many algorithms have been investigated in order to obtain high quality images in CT. However, for their practical application to the CT system, it is necessary to solve many problems with the scanned data including patient motion, data truncation, photon noise and circuit noise. In this paper, we describe practical aspects focusing on the background of image reconstruction techniques generally used on the CT system, initiatives aimed at a practical use of iterative reconstruction, schemes for effective application, along with problems to be raised in applying the reconstruction on the CT system and its corrections.

View full abstract

This paper provides an overview of the compressed sensing approach and its contribution in the development of reconstruction algorithms in x-ray computed tomography. We briefly review the recent advances in the topic and refer to the related applications. Finally, we discuss the challenges related to the development of the compressed sensing based reconstruction algorithms and possibilities to be used in clinical scanners.

View full abstract

We describe practical aspects of iterative image reconstruction methods currently used in emission CT. In particular, we report on accurate modeling method of the system and present state of the image reconstruction methods using GPU as a solution for the increase of eventual computation amount. In addition, we introduce state-of-the-art multi-modality systems such as PET-CT and PET-MRI, and the TOF-PET systems.

View full abstract

Statistically-based iterative image reconstruction is widely used in emission computed tomography. This paper explains the mathematical foundations and computational techniques in statistical image reconstruction through the formulation and solution of the maximum-likelihood (ML) estimation problem, and introduces some pioneering techniques as extensions of the ML estimation.

View full abstract

MRI uses gradient magnetic field to add position information of spins on MR signals as modulation of frequency and/or phase. Observed MR signal is Fourier transformed from measurement space (k space) to real space (image). Before Fourier transform, processing such as base line correction, gridding are performed. After Fourier transform, processing such as distortion correction, shading correction are performed. Parallel imaging that uses characteristics of multiple RF receiving coils, are utilized to fast data acquisition.

View full abstract

Due to the increase in the size of medical images, huge computational complexity is required for image analysis and feature extraction. In order to address this problem, parallelization of processing has become increasingly important in recent years. The Message Passing Interface (MPI) is a common library for developing parallel programs. High-speed processing can be realized using MPI, but it is necessary for programmers to consider how to split the data and how to process the data transmitted and received between CPUs. Programming therefore becomes more complicated. In previous research, a parallel programming interface based on MPI was developed, but the method employed in this previous research can only split simple arrays. In the field of medical imaging, the Insight Segmentation and Registration Toolkit (ITK) is the most commonly used image-processing library worldwide. However, because ITK has a unique data structure, the methods employed in the previous research are not applicable to ITK. The authors have therefore developed a set of macros based on MPI that can be applied to programs using ITK. The authors have also verified the effectiveness of the proposed method based on experimental results.

View full abstract

As the amount of information contained in 3-dimensional medical images increases, the computational cost of computer-aided diagnosis (CAD) becomes higher. The calculations must therefore be performed faster. The authors have focused on parallel computing methods using the Compute Unified Device Architecture (CUDA) as one of the techniques for general-purpose computing on graphics processing units (GPGPU), which is a technology for diverting the computing resources of GPUs to perform general-purpose computations. In order to use CUDA effectively, programmers must carefully consider the various requirements of parallel programming and CUDA. The Insight Segmentation and Registration Toolkit (ITK) is widely used as a standard medical image processing library in the field of medical image processing. However, because ITK has a unique data structure, it cannot be used in programs with CUDA as is. To address these issues, the authors propose a method for developing parallel ITK programs with CUDA by means of sequential ITK programs. The authors have conducted comparative evaluations of computational speed and usability, and the results confirm the usefulness of the proposed method.

View full abstract

Image registration for ultrasound images acquired at different stages of diagnosis and treatment is extremely important. Reliable image registration facilitates the monitoring of tumors over time and the evaluation of therapeutic effectiveness, which are essential for determining the most appropriate course of treatment. Aggressive research has led to considerable progress in image registration methods for CT, MR, and PET. However, due to various factors peculiar to the ultrasound modality, there have been significantly fewer image registration research studies on ultrasound than other modalities. Due to the mutual information calculation algorithm, the increases (or decreases) in mutual information of sub-regions do not necessarily reflect the increase (or decrease) in mutual information of the combined entire region. To address this point, we previously proposed an effective ultrasound image registration method employing the unified mutual information derived from a few multiple small regions. However, these regions were set in morphological information-rich areas by operator action. In order to improve consistency and reproducibility, we now propose an automatic VOI placement method in which a comparatively large number of VOIs are set at the center of the image. We have evaluated the effectiveness of the proposed method using clinical data and have confirmed that it can provide the same level of registration performance as manual VOI setting.

View full abstract

In this paper, we describe abdominal lymph node detection methods intended to assist in preoperative diagnosis before surgery for colon or stomach cancer. These methods are able to provide assistance in determining the area to be resected at the time of preoperative diagnosis and to display the detected lymph nodes during the surgical procedure by detecting the lymph nodes in which there is a high likelihood of metastasis. When the previous method is employed, many false positives are also detected. We have therefore improved the blob-like structure enhancement processing for detecting lymph nodes and the false positive reduction procedure based on feature analysis for increasing the detection accuracy. The improved method was applied to 28 abdominal CT images, and the results showed that it was possible to detect 65 of 95 lymph nodes with about 13 false positives per case.

View full abstract

We have developed a system to accelerate the snake algorithm using the Cell Broadband Engine (Cell processor). The snake method is a representative technique for the extraction of outlines such as the contours of organs in medical imaging diagnosis. However, it is difficult to set the optimal parameters for the snake method automatically, and it is therefore often necessary to perform parameter setting repeatedly in order to obtain the desired image. In our parallel computing system using Cell processors, we avoid this problem by assigning each processor a different parameter, allowing us to calculate multiple parameters at the same time for outline extraction by the snake method. The desired image is obtained in many of the simulations. The Cell processor is a representative multicore processor. Our system includes 16 PlayStation 3 consoles (home video game consoles manufactured by Sony), and each PlayStation 3 contains a Cell processor. Since we are able to employ the six calculation units in each Cell processor, our system with 16 PlayStation 3 consoles can perform 96 outline extraction calculations with different parameters at one time.

View full abstract