The Pharmaceuticals and Medical Devices Agency (PMDA) approved a computer assisted diagnostic program for cerebral aneurysms from magnetic resonance angiography (MRA), co-developed by the Department of Diagnostic and Interventional Radiology Osaka City University Graduate School of Medicine, as the first medical device using an artificial intelligence (AI) in Japan on September 17, 2019. This is a dawn of the AI era in medicine. Research in the AI of radiology can be broadly classified into four categories : object classification, object detection, object segmentation, and image processing. In this article, these categories were introduced with studies in our department and roles of medical practitioners in the AI era were discussed through an experiment of a difference between AI recognition and medical practitioner recognition.

In this study, we evaluated the effect of a newly developed super-resolution processing using Deep Convolutional Neural Network (DCNN) on the images processed with the conventional dynamic processing in digital radiography (DR). We used human phantom images to obtain case samples of which each image included a lateral view of thoracolumbar junction. All case samples were processed without and with dynamic processing, which were ranging from 1 to 4 steps of image enhancement processing. Deep Denoising Super Resolution (DDSR) were trained and assessed using the supervised image (the original image) and the low-resolution image degraded to 1/3 of matrix size by binning from the original image. The effects of the DDSR on the conventional dynamic processing were assessed using the modified Ura method of Scheffe's paired comparison. The average psychological measures for interpreting vertebral body of thoracolumbar junction tended to increase as the enhancement of dynamic processing increased, regardless of the application of DDSR. The correlation between the average psychological measures without and with DDSR was very high with a correlation coefficient of 0.98. We conclude that the effect of the DDSR on the conventional dynamic processing would be neglected on the observation of a diagnosis object in DR.

With the progress of post-genomic research, the relationship between various tumors and genes has been elucidated. However, in the field of radiology, there is not much research to understand the molecular and genetic backgrounds involved in the image phenotype of a lesion. The purpose of this study is to develop image data mining technology to analyze the relationship between image phenotype and genotype of a lesion. Fat-suppressed T1-weighted images of 49 cases were collected from TCGA-BRCA (The Cancer Genome Atlas Breast Invasive Carcinoma) public database. The slice with the largest tumor diameter was selected from the MRI and the tumor regions were manually segmented. A total 371 radiomic features including size, shape, texture, etc. were calculated from the tumor region. By using CART (Classification and Regression Trees) algorithm with these radiomic features as input data, a classification tree that outputs 5 breast cancer subtypes were automatically generated. The overall accuracy of the classification tree for identifying 5 breast cancers was 83.7% (41/49). By applying the proposed method, it is possible to visualize the relationship between image phenotypes and breast cancer subtypes.

Volumetric breast density (VBD) is considered to be the most reliable measure among those that evaluate breast density (BD) using mammographic apparatus, but its theory relies solely on primary x-rays that penetrate breast tissue. Thus the purpose of this study is to evaluate scatter x-ray influence on VBD, since the amount of scatter x-rays can not be ignored in usual mammographic exposures. Using breast equivalent phantoms whose BD is 0%, 50% and 100% respectively, we first measured scatter-to-primary ratio (SPR) with the lead disk method by changing kVp, phantom thickness, grid (+–) and BD. The true SPR was derived from 3 mm diameter lead disk data by utilizing extrapolation towards 0 mm diameter. Through the measurement, we found that SPR slightly increased as BD increased, while it vastly depended on other factors. Based on these measurements, we calculated the error that was caused by the scatter x-rays on the BD estimated. As a result, the error stayed within 1% for the most cases, though it reached as large as 10.7% at the extreme case of 8 cm phantom thickness (100% BD) without anti-scatter grid. These results are expected to contribute to efforts in improving the accuracy of VBD.