This paper proposes a method for disclosing the reasoning behind computer-aided diagnosis (CADx) based on a Bayesian network. The purpose of this method is to promote the acceptance of CADx by physicians by providing the reasoning behind the inferences. The proposed method first calculates the influence ratio to the inference result for each subset of input information. It then selects some subsets that have large influence ratios and shows them as the reasoning or grounds for the inference. In experiments using artificial data with known classification rules, the proposed method detected correct rules for about 90% of the data. With regard to clinical data, the average value for the effectiveness of reasoning as judged by two physicians was 3.4. This value is greater than “3,” which is considered a reasonable grade.

View full abstract

We propose a novel image denoising strategy based on the correlation in the FREBAS transformed domain. FREBAS transform is a kind of multi-resolution image analysis which consists of two different Fresnel transforms. It can decompose images into down-scaled images of the same size with a different frequency bandwidth. Since these decomposed images have similar distributions for the same directions from the center of the FREBAS domain, even when the FREBAS signal is hidden by noise in the case of a low-SNR image, the signal distribution can be estimated using the distribution of the FREBAS signal located near the position of interest. We have developed a collaborative Wiener filter in the FREBAS transformed domain which implements collaboration of the standard deviation of the position of interest and that of analogous positions. The experimental results demonstrated that the proposed algorithm improves the SNR in terms of both the total SNR and the SNR at the edges of images.

View full abstract

In the field of dentistry, an integrated display environment of a 3D dental cast and panoramic radiography can provide a clear understanding of the patient's dental condition. Not only visual but also haptic display appears to be effective in order to present an integrated image. This research proposes a cross-reference system that enables users to indicate a point on the main image and obtain additional information from the other image through haptic feedback. In the proposed method, a virtual panoramic image created from a 3D dental cast and a real panoramic image are registrated based on the dental contour to obtain integrated multimodal images. The proposed method was evaluated using manually collected points in each image. The results indicated that a registration error of 1.4 mm was acceptable for obtaining a sufficiently clear understanding of the dental condition. Furthermore, vibrotactile and haptic sensation could be presented according to the interior and the shape information of the dental image.

View full abstract

This report presents a novel method for matching small pulmonary nodules (SPNs) that is robust to the temporal changes in the nodules on follow-up thoracic CT images. The observation of SPNs on follow-up CT images is very important in the diagnosis of lung cancer, but such observation places a heavy burden on the physician. This is because the differences in the position of the patient on the CT table and the effects of cardiac and respiratory motion usually lead to deformation on follow-up CT images. In addition, due to the development of metastases or the effects of chemotherapy, nodules may appear or disappear or may fuse or separate. In order to reduce the burden on the physician, we have developed a method for assisting him or her during the observation of SPNs on follow-up CT scans. The proposed method employs both rigid and non-rigid registration methods to co-register follow-up CT scans in order to compensate for lung deformation. Taking into consideration the temporal changes in the SPNs, we calculate three matching patterns to identify the corresponding nodules based on various features such as the location of the center of gravity, the diameter, and the average intensity. We applied the proposed method to 3 patients (14 thoracic CT scans) and correctly matched 94.6% of the segmented SPNs. The results also showed that our method is robust to changes in SPNs such as the appearance/disappearance or the fusion/separation of nodules.

View full abstract

The registration of 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 treatment effectiveness, which are essential for deciding on the most appropriate therapeutic plan. Aggressive research has led to considerable progress in image registration for CT, MR, and PET. However, due to various factors peculiar to the ultrasound modality, there have been far fewer research studies on image registration focusing on ultrasound than on other modalities. Recently, diagnostic ultrasound systems that can acquire and record three-dimensional volume data continuously have been developed. In ultrasound diagnostics, it is extremely important to be able to reliably match positions for a series of images obtained at various stages of the diagnostic and treatment processes. This has led to strong demand for a reliable image registration technique. We have proposed a unique registration technique using volume similarity measures based on multi-small regional mutual information and have demonstrated its effectiveness. In order to apply this technique in the clinical setting, it is necessary to improve its processing speed. The present study introduces a computational acceleration method for our registration technique which employs a graphics processing unit (GPU). The results indicate that it may be possible to increase the processing speed by a factor of 30 or more.

View full abstract

It is important to estimate rib cage motion quantitatively because lung deformation is caused by motion of the rib cage and diaphragm. In this paper, a patient-specific model of rib cage motion between inhalation and exhalation is proposed based on anatomical knowledge and rib cage kinematics. Time-series geometric changes in the patient-specific 3D rib cage configuration (including the ribs, spine, and sternum), which is segmented based on exhalation CT images, can be simulated using the proposed method. In addition, the validity of the proposed rib cage motion model was evaluated using inhalation and exhalation CT images.

View full abstract

DRAMA (Tanaka E, Kudo H, Phys Med Biol, 2003) provides fast convergence with a reasonable signal-to-noise ratio by updating the image with row-action by applying the subset-dependent relaxation parameter. A new method that automatically determines the next relaxation parameter based on the previous image update was proposed by Kudo in 2008. This method controls the relaxation parameter by monitoring the image convergence phase and is applicable to various conditions of the subset, access, and control unit of the relaxation parameter. In this study, we implemented this method on a whole-body 3D PET scanner (SET-3000G/X, Shimadzu Corporation) and investigated the optimal access order for subsets and the control unit of the relaxation parameter by comparing contrast-noise performance in NEMA IEC Body Phantom images. The performance obtained with DRAMA (one iteration) was significantly superior to that obtained with MLEM (multiple iterations). The optimal access order was the parallel sinogram first, followed by the remaining oblique sinograms in descending order. As for the control unit of the relaxation parameter, image convergence was expected to be faster with a smaller control unit, but its behavior was unstable due to statistical noise. The new method that controls the relaxation parameter in DRAMA was found to work well and to provide realistic 3D PET data. We will investigate the behavior of convergence values with various activity distribution conditions and statistical noise levels using clinical data in our future work.

View full abstract

[in Japanese]

View full abstract