This article was prepared based on a special lecture given at the meeting of the Society of Medical Image Information on June 2, 2018, which was a summarized version of the lecture provided to students at Doctoral level in Graduate School of Gunma Prefectural College of Health Sciences. The contents include（1）Why do we need to nurture young researchers, and how to guide them; （2）What is a leadership in science and research, and how to achieve the leadership; （3）How to manage and administer research groups and/or large organizations; （4）Why do we need to publish articles in English, and how to publish it.
This paper presents a brief overview of machine learning and its application to medical images. Machine learning is a technology or technique that has been spreading rapidly due to the development of computer environment and the enormous amount of data available. Deep Neural Network （DNN）, which is one of machine learning methods, has been studied explosively in recent years. This paper introduces the research applying DNN to medical images.
Image-guided radiotherapy （IGRT） systems using kilovolt cone-beam computed tomography （kV-CBCT） images are being commonly used for highly accurate patient positioning in lung stereotactic body radiotherapy （SBRT）. However, current IGRT procedures are based on bone structure and subjective correction. Therefore, the purpose of this study was to investigate an automated and robust estimation framework for lung tumor location in kV-CBCT images to improve target-based patient positioning for lung SBRT. One-hundred-and-sixty kV-CBCT images from 40 clinical cases treated with SBRT were used. The proposed framework comprised four steps, i.e., determination of a search region,extraction of a tumor template, preprocessing for enhancement of the tumor region, and estimation of the tumor location by a template-matching technique. Original, edge enhancement, and tumor enhancement images were obtained by enhancement of a tumor region based on each form of preprocessing and were used for template matching. The mean Euclidean distances of location errors for original, edge enhancement, and tumor enhancement images were 1.2 ± 0.7 mm, 5.5 ± 10.1 mm and 2.7 ± 4.4 mm, respectively. These findings suggested that the proposed automated framework may be robust for estimating the location of lung tumors in kV-CBCT images for lung SBRT.