Magnetic resonance imaging (MRI) T1-weighted images are often used in the diagnosis of diseases such as brain degeneration because MRI has superior contrast of soft tissue. Although the clinical use of MRI has traditionally been limited to visual evaluation by radiologists, but recent advances in image analysis techniques possible to automatically quantitatively evaluate brain volume voxel by voxel and within anatomical regions of interest. This has led to many reported applications in machine learning using a hues amount of MRI data sets. However, medical imaging, a type of clinical data, involves various problems caused by their clinical backgrounds and imaging scanner. In this paper, we focus on dementia and structural MRI, and introduce the analysis methods and their applications to machine learning, as well as the problems that exist in machine learning using medical images.

In the field of imaging diagnosis of dementia, the concept of imaging biomarkers has proposed the research criteria of ATN classification corresponding to amyloid （A）, tau （T）, and neurodegeneration （N）. The use of amyloid PET （A）, tau PET （T）, and FDG cerebral glucose metabolism PET （N） as imaging biomarkers will enable ATN classifica tion and is expected to contribute to the differential diagnosis of dementia and the elucidation of the progression process of Alzheimerʼs disease.

Dementia with Lewy bodies (DLB) is highly prevalent dementia disorder in elder population, after Alzheimerʼs disease. DLB has multiple involvement of neurotransmitter systems with non-specific structural changes. Clinical symptoms as well as SPECT imaging is effective diagnostic marker which is incorporated in the diagnostic criteria of DLB. Since DLB is treatable, early recognition and diagnosis is pivotal in clinical practice. In this review dopamine transporter image, cardiac ¹²³I-metaiodobenzylguanidine (¹²³I-MIBG) image, and perfusion image, are detailed, in terms of scan protocol, image analyses, a role as a diagnostic tool, and clinical correlates of DLB. Reduced striatal dopamine transporter binding, reduced cardiac ¹²³I-MIBG binding, and diffuse cortical hypoperfusion with most prominent occipital lobe involvement is the typical imaging finding of DLB. Pearls and pitfall for diagnosing DLB is discussed.

MRI is essential for the early and differential diagnosis of Alzheimerʼs disease and for the evaluation of its progression. Three-dimensional T1-weighted images can be acquired easily in routine clinical practice, and have recently been applied to network analysis. In recent years, T1-weighted images have also been applied to network analysis. In addition, a method that focuses on the similarity of gray matter has made it possible to detect network abnormalities at the individual level. These new methods contribute to the understanding of the pathogenesis of Alzheimerʼs disease by allowing us to examine the effects of amyloid-𝛽 and tau protein accumulation on white matter connectivity and brain networks.

The edge spread function (ESF) in the circular edge method is obtained directly using the reconstructed circular disk, but line spread function (LSF) is determined by differentiating the ESF. We investigated the ESF influenced by subpixel sampling for the filtered backprojection (FBP) images by a simulation study, and estimated the full width at half maximum (FWHM) using curve fitting to ESF assuming the LSF is given as Gaussian function. A phantom consisted of 184 mm background disk and 20 mm signal disk; the contrast was 15%, 7%, and −7%; a 200-mm field of view; image size was 1024 × 1024 matrix and 512 × 512 matrix. Edge of phantom image was blurred by the cumulative distribution function (CDF) of Gaussian function with a known FWHM, and parallel beam projection data with 1024 (512) radial bins and 1024 (512) /180° angular views were reconstructed. The circular edge method is affected by subpixel sampling, differentiation and noise, and CDF Gaussian fitting is effective in reducing the ESF fluctuation. The FWHM by an approximation expression agreed well with that of numerical experiment when FWHM > 0.276 mm for 1024 × 1024 images and FWHM > 0.677 mm for 512 × 512 images.

Iterative image reconstruction using nonlinear sparsifying transform with L1 regularization for CT has been reported to preserve contour and texture while decreasing statistical noise. We investigated the spatial resolution characteristics of this algorithm when applied to a conventional angular views with simulation study. A two-dimensional numerical phantom with a 512 × 512 matrix consists of 184 mm background disk with various 20 mm diameter contrast inserts (−60% to +60%). An ideal parallel beam projection data without degrading factors such as scatter and beam hardening was assumed. The spatial resolution expressed as FWHM was estimated from the edge spread function for disk. The FWHM of L1 reconstruction using nonlocal means filter (NLM) and bilateral filter was highly dependent on the contrast, while median filter was not. A threshold 𝛿* exists in parameter 𝛿 of L1 reconstruction with NLM (L1-NLM) for each object with different contrast; 𝛿* determines the spatial resolution blurring contour most. Contour preservation works effectively if 𝛿 is set sufficiently smaller than 𝛿* of each object, but the contour is remarkably blurred if it is set to larger than or equal to 𝛿*. This causes the contrast dependence of spatial resolution in L1-NLM.

It is well known that tissue stiffness (elasticity) changes from the early stages in many diseases such as cancer tumors, arteriosclerosis, and liver cirrhosis. For this reason, if the distribution of tissue elasticity can be visualized in addition to tissue morphology and functional information such as blood flow by conventional medical images, namely CT, MRI, and ultrasound, it will be useful for early diagnosis, differential diagnosis of benign and malignant, and evaluation of therapeutic effects such as chemotherapy. To meet these clinical needs, we promoted research and development of ultrasonic elastography through industry-academia-government collaboration, and in 2003 achieved the practical application as the worldʼs first clinical device. In addition, we have developed diagnostic guidelines for the dissemination of new technologies to clinical practice. As a result, many ultrasound devices now have an elastography function, and are widely used in clinical fields as a new diagnostic imaging method for breast cancer, chronic hepatitis, and so on.

This paper introduces the development of an endoscopic diagnosis assistance system using artificial intelligence technology, which has received JAMIT Technological Achievement Award. This system detects colonic polyp from colonoscopic video in real-time and performs pathological classification of a detected colonic polyp. This system implements the following function: (1) colonic polyp finding assistance, (2) colonic polyp classification into neoplastic or non-neoplastic categories, and (3) detailed classification of neoplastic polyps. Especially the EndoBrain system, which has the function of classifying colonic polyps into neoplastic and neo-plastic categories based on ultra-magnified endoscopic images, is the first medical device using AI/ML technology obtained a medical device certificate under governmental regulation. This paper shows an overview of AI-based assistance systems for colonoscopy and perspectives on this field.

In Part 2, we will explain the method of reconstructing a refraction-contrast CT image from a refraction-contrast image obtained by X-ray Dark-Field Imaging (XDFI). A refraction-contrast image is a brightness representation of the refraction angle of X-rays propagating through a subject, whereas a refraction-contrast CT represents the refractive index distribution of the subject. Generally, obtaining a CT requires a line integral of the subject as a projection, and to reconstruct the refractive index distribution, the refractive angle must be converted to a line integral of the refractive index distribution. In this course, three projection equations that express the relationship between refraction angle and line integral obtained in the past will be introduced, and the CT reconstruction method for each will be explained.