Yoshiharu Hirata

Comparison of increased venous contrast in ischemic stroke using phase-sensitive MR imaging with perfusion changes on flow-sensitive alternating inversion recovery at 3 Tesla:

Background Increased venous contrast in ischemic stroke using susceptibility-weighted imaging has been widely reported, although few reports have compared increased venous contrast areas with perfusion change areas. Purpose To compare venous contrast on phase-sensitive MR images (PSI) with perfusion change on flow-sensitive alternating inversion recovery (FAIR) images, and to discuss the clinical use of PSI in ischemic stroke. Material and Methods Thirty patients with clinically suspected acute infarction of the middle cerebral artery (MCA) territory within 7 days of onset were evaluated. Phase-sensitive imaging (PSI), flow-sensitive alternating inversion recovery (FAIR), diffusion-weighted imaging (DWI) and magnetic resonance angiography (MRA) were obtained using 3 Tesla scanner. Two neuroradiologists independently reviewed the MR images, as well as the PSI, DWI, and FAIR images. They were blinded to the clinical data and to each others findings. The abnormal area of each image was ultimately identified after both neuroradiologists reached consensus. We analyzed areas of increased venous contrast on PSI, perfusion changes on FAIR images and signal changes on DWI for each case. Results Venous contrast increased on PSI and hypoperfusion was evident on FAIR images from 22 of the 30 patients (73%). The distribution of the increased venous contrast was the same as that of the hypoperfused areas on FAIR images in 16 of these 22. The extent of these lesions was larger than that of lesions visualized by on DWI in 18 of the 22 patients. Hypointense signals reflecting hemorrhage and no increased venous contrast on PSI and hyperperfusion on FAIR images were found in six of the remaining eight patients (20%). Findings on PSI were normal and hypoperfusion areas were absent on FAIR images of two patients (7%). Conclusion Increased venous contrast on PSI might serve as an index of misery perfusion and provide useful information.

Effect of pre-processing on performance of a neural network with one-dimensional sampling from X-ray images of chest

To judge whether the images from X-ray of chest are abnormal or not, we constructed the system including a three-layered neural network with one-dimensional sampling and investigated the effect of pre-processing, FFT (Fast Fourier Transform) on the performance of the system for medical diagnosis support. We used one-dimensional lines from two-dimensional images from X-ray of chest as input patterns for the neural network. From the results, we found that square error of our neural network decreased for learning patterns, and that there was broden region, where the FAR(False Acceptance Rate) and FRR(False Rejection Rate) were zero. This indicated that our system with the combination of FFT and the one-dimensional sampling was useful for judgement whether the images from X-ray of chest were abnormal or not. In addition, we found that the network with the less learning iterations for convergence did not miss for any abnormal test patterns with low FRR for normal test patterns. Thus, we suggested a possibility of useful medical diagnosis support system by using the neural network.

Automatic restoration of X-ray Photography by estimating point spread function

In this paper, an image restoration method using Iterative Blind Deconvolution is proposed. It removes a blur by estimating the Point Spread Function (PSF) of Chest X-ray Photography (Chest-XP) automatically. The Chest-XP contains a blur in the principle of taking images. The conventional methods remove a blur by measuring a phantom image or using the blind deconvolution. However, there are problems that a phantom image should be precisely taken, or the parameters of evaluation function are empirically and manually given. In the proposed method, they are automatically decided by using edges and noises in Chest-XP and estimated images. We demonstrate the effectiveness for interpolating Chest-XP and camera images.

Resolution improvement of point dose distribution in intensity modulated radiation therapy

In intensity modulated radiation therapy (IMRT), there are verification methods for the point dose distribution, such as two dimensional (2D) detector or film based verification. The 2D detector based verification is known to be easier than the film based one, however the obtained point dose distribution is lower quality. In this paper, we apply a super-resolution method to improve the image quality of point dose distribution obtained from the 2D detector. We conducted an experiment to evaluate the super-resolution images and its ideal reference images. Each dose distribution obtained by proposed method satisfied the criteria in gamma analysis. Although the failed point is slightly observed, clinical problems are probably not caused. Thus the feasibility of the proposed method is confirmed.

Construction of high-performance systems with neural networks for detection of abnormal areas from chest X-ray images

We constructed systems with neural networks using one-dimensional numeric sequences from chest X-ray images for detection of abnormal areas and investigated the effect of pre-processing for input patterns on performance of the systems. We changed the number of data in one-dimensional numeric sequences, and applied differential filters to the sequences. Then, we produced the input patterns which consisted of pixel values, 1st-order differential coefficients and 2nd-order differential coefficients. From the results, we found that the performance of the systems using the input patterns of pixel values was superior to those using the others. The best value of Az was 0.95 when the number of units in input layer in the systems was 16. The performance of the systems in this study is thought to be comparable with that of the systems using two-dimensional areas in the images with complex image processing.

Experience of a small PACS in Tottori University Hospital

Picture archiving and communications systems (PACS) are expected to become feasible as the way of total radiologic image management. As the first step towards computerized image management, we installed an experimental small PACS with a digital optical archive for magnetic resonance and digitized radiographic film images. During the first 10 months of operation, problems were caused by the frequent mechanical troubles on the acquisition devices, impractical schemes for transmission, and troublesome retrieval and display of images. Although some of them have been solved by the improvements of both hardware and software, the mechanical instability of the digitizer and the inadequate operating procedure of the workstation still remain. As to the characteristics of PACS, the capability of access to every achieved radiologic image is the great advantage for us because the image integration has been impossible in the form of radiographic films for the traditional image management system of our hospital. However, the complicated image display procedure with high overheads is far from ideal for radiologists. The efficiency of image display must be increased to the level of a conventional film viewing system. Furthermore, research is needed on the small PACS to get consensus for the acceptance of a more widely implemented system.

Effect of number of input layer units on performance of neural network systems for detection of abnormal areas from X-ray images of chest

We constructed neural network systems using one-dimensional numeric sequences from X-ray images of chest for detection of abnormal areas in the images and investigated the effect of number of input layer units on performance of the systems. In order to construct the neural networks with different number of input layer units, we changed the number of data in the input patterns, which were one-dimensional numeric sequences obtained from the two-dimensional images, by using averaging filters. Then, we produced the input patterns which consisted of 16, 32, 64 and 128 numbers of data from the one-dimensional numeric sequences. From the results, we found that the size of detectable abnormal areas in the systems was dependent on the number of input layer units in the range from 16 to 128. In addition, the performance of the systems using one-dimensional numeric sequences as the input patterns was comparable with that of the systems using two-dimensional areas. Therefore, the system used in this study is thought to be useful for the detection of abnormal areas from X-ray images of chest.