Compressed sensing (CS) is attracting growing concerns in sparse-view CT image reconstruction. The most explored case of CS is total variation (TV) minimization. However, images reconstructed by TV usually suffer from some distortion, such as patchy artifacts, improper serrate edges and loss of image textures, especially in practical CT images. Most existing CS approaches including TV achieve image quality improvement by linear transform to object image. Considering the success of nonlinear filters in image processing such as denoising, we propose to replace linear transform with nonlinear ones in CS on sparse-view reconstructions as to obtain further promotion. Median filter, bilateral filter and nonlocal means filter were respectively explored and combined in CS framework. As the iterative method, majorization-minimization (MM) based iterative-thresholding (IT) method was utilized. Experimental results with both digital and clinical images consistently demonstrated that nonlinear filter based CS has potentials in achieving further image quality improvements compared with typical TV minimization.
The bloodstream system of the pulmonary field consists of two systems: pulmonary arteries pass and aorta pass. Until now, the influence of recirculation has been excluded by fitting a gamma function, because the recirculation from aorta is less than 10 percent of the blood flow from pulmonary arteries. In this study, we consider the way to estimate influence of recirculation by modeling the impulse response of a blood flow system that consists of 2-inputs with pulmonary arteries pass and aorta pass firstly, and we propose the way to analyze two blood flow systems separately. An experimental prototype system using these methods was applied to three cases with pulmonary tumor. Furthermore, we compared the global blood volume of the whole pulmonary field inferred from the signal intensity of the pulmonary arteries and the left atrial, with the total of the local blood volume of the whole pulmonary, by using perfusion MRI. The validity was considered by comparing the percentage of the recirculation which occupies a pulmonary field with the cardiac output volume calculated using Fast GE cine MRI.
We have developed a guided image filtering (GIF) using a noise model (NM) of whole-body positron emission tomography (PET) images for cancer screening studies. The noise standard deviation in each voxel was derived from a noise image (NI). The NI was created by taking the difference of two reconstruction images created from two group data. The list-mode PET data were divided into odd and even group data. The NM was created by taking the mean of distribution expressed the relation between voxel values and noise standard deviations in the pair of the each PET image and NI for 185 samples. We compared the image quality between the Gaussian filtering (Gauss) and GIF with/without the NM using the whole-body PET image embedded an artificial hot spot into lung, kidney and bladder. We also tested the GIF with the NM using the whole-body PET image mismatching to the NM. The GIF without the NM produced lower contrast recovery coefficient than the Gauss in the case of a hot spot with 2:1 contrast inserted in the lung; however, this was prevented by GIF with the NM.
One of the most important factors for research and development of computer aided detection/diagnosis is an integrated large-scale database which has a link between the medical image and many types of incidental information. We built an integrated lung nodule database which contains 1240 cases. In this database, all cases include the following data: (1) thin slice CT image data; (2) the position information of the lung nodule; (3) structured imaging findings given by consensus of two board-certified radiologists; (4) the confirmed diagnosis; (5) clinical information; and (6), the diagnosis reasoning of the lung nodule given by board-certified radiologists. In this study, we introduce the details of this database and the efforts of the construction.
As a discrimination method for estimating invasive potentiality and prognosis of oral squamous cell carcinoma (OSCC), the mode of invasion (Yamamoto Kohama-criteria) centered on the form of pathological tissue specimen has been known to be clinically useful, especially in Japan where it is used frequently. However, evaluation of this mode of invasion is based on subjective visual observation, which has created a large gap between evaluators and facilities using such the mode of invasion. This problem is such that unification of objective evaluations is a challenge. Therefore, this study aimed to develop a method of automatically determining the mode of invasion by medical image processing using digital images of the invasion front of OSCC. A shaped feature of the invasive fronts was extracted to create a classifier by Random Forest, a machine learning algorithm, in conjunction with the mode of invasion opted for based on the judgment of the clinician. As a result of inputting multiple test images to the created classifier, it was confirmed that the classifier outputs a decision that is very close to the judgment of the clinician. Therefore, automatic determination of the mode of invasion by a medical diagnostic imaging system was shown to be feasible and of high accuracy.
In the parts (1) and (2) of this tutorial, PET detectors and its signal processing for data acquisition were reviewed. In order to generate accurate images, it is necessary to apply various corrections for PET measurement data containing many kinds of noise components and affected by attenuation of body. In this paper, I explain the reconstruction method to generate PET images by processing the measurement data and applying the various kinds of correction methods.