In diagnostic imaging, a radiologist confirms the purpose of examination and the type of image by looking at a request form and medical records from the attending physician. In interpretation, the specialist judges the presence or absence of abnormal signs to identifies and describes them. Furthermore, the specialist adds interpretation of the findings, pathological analysis, and differential diagnosis to reach the final diagnosis. Generally, the specialist analyzes different kinds of images to identify the site, morphology, margin, boundary, properties, etc. of the abnormal signs. At present, it is difficult for AI to handle different types of images at the same time.

Comparison with normal anatomy and past images is particularly important for discovery of abnormal findings. However, normal anatomy is significantly different among ages, genders, individual differences, and imaging equipment. Thus, it is necessary to collect supervising image data in each category for AI learning. Therefore, AI development in diagnostic imaging indispensably requires sufficient knowledge of such normal anatomy as well as imaging methods. Nevertheless, AI is also effective for assistance of filming, denoising, etc. and is expected to be incredibly useful in areas that are facing a shortage of specialists in diagnostic imaging.

X-rays and CT are typical medical devices for diagnostic imaging. The used X-rays are characteristic as short wavelength and high transparency. This transmission effect is also determined by the density and length of the object material. Image is created by transmission, photosensitivity, and ionization of X-rays. Transmission of substances by X-rays provides energy to atoms and molecules which will emit electrons. Then, X-rays as light will be absorbed by electrons having energy. Finally, the shade of the image will be affected by the number of X-rays that remains unabsorbed.

This shade is affected by the type and state of tissues and visualized by this contrast and tangent line formation. At present, this process is performed and visualized by computer. CT is a three-dimensional type of X-rays but uses the same principle for imaging. However, CT takes the standard that the CT value of water is 0. CT value has 2000 levels of shade but can only be expressed by 256 levels. Thus, the range of expression must be selected according to each application. Some cases require the use of contrast media. Furthermore, the shade is influenced by both volume and density of the objects.

MRI is a device that photographs various objects that receive radio waves with a certain frequency in a magnetic field. MRI images the state of protons at a certain time by utilizing the difference of the density of hydrogen nuclei (protons) contained per volume. Radio waves of a specific frequency make hydrogen nuclei in the object face in the same direction (magnetic resonance or excitation) and turning off the radio waves will lead them back to the original direction (relaxation). The contained number of protons is different among tissues thus, the relaxation speed is also different. This difference can contribute to imaging each tissue.

Generally, MRI produces T1-weighted and T2-weighted images, which are respectively created through vertical and horizontal relaxations. The relaxation speed in each tissue is different between T1 and T2-weighted images. Thus, the proper use of both is important. Furthermore, the situation of each tissue can be understood in more detail by Fluid Attenuated Inversion Recovery (FLAIR), which inhibits influence of water, Diffusion-Weighted Image (DWI), which images the degree of diffusion of water molecules, and Susceptibility-Weighted Image (SWI), which emphasizes the difference of magnetic susceptibility of tissues.

PET (Positron Emission Tomography) is a device that measures the annihilation radiation created at extinguishment of positrons and visualizes an image of tissues and organs. PET quantitatively images information as to physiological function, particularly blood flow and metabolism. PET examination uses glucose and the like with positron emitting nuclides. These positrons are generated and emitted from an unstable nucleus having excessive protons and collide with nearby electrons and extinguish to release a certain amount of energy in the end. This energy (γ-rays) is converted into an optical signal by scintillator and amplified by photomultiplier tube, then its position information will be acquired by position calculation circuit. However, various errors occur depending on the filming conditions, thus adjustment of the errors is indispensable.

The measured amount of radioactivity at a given location is reproduced into image by the Filtered Back Projection (FBP) and evaluated by the Standardize Uptake Value (SUV). The average SUV amounts to 1. Since PET, which provides functional images, is different from morphological examinations like CT, it can visualize inflammation and infection as well as effectively predict therapeutic effects and prognosis. Furthermore, it is expected that AI can improve diagnosis and prediction of pathological conditions as well as image production by PET.

Ultrasonography irradiates high-frequency sound waves of 20,000 Hz or more and produces images from the reflected sound waves. It utilizes the ability of ultrasonic waves to decompose the space, contrast, and the time. Contrast resolution is the ability to describe the differences of acoustic impedance of tissues, which can provide images that are more sensitive than X-ray CT images. The advantages of ultrasonography are non-invasiveness, convenience, and real-time display. The drawbacks are the difficulty of operation and the limitations on display by gas, fat, and the distance.

IVR is an abbreviation for interventional radiology and means performance of real-time treatment at the same as diagnostic imaging, such as X-ray fluoroscopy, angiography, ultrasound, and CT. It has the advantage of real-time observation of detailed images, but also has the disadvantages of exposure to radiation and invasiveness. The performance of IVR includes embolization for cancer, aneurysm, and bleeding, vasodilation for stenosis, angiography for superior vena cava syndrome, ablation for cancer, implantation for catheter placement, biopsy for lesion diagnosis, drainage for excretion and so forth.

The computer world only has “0” and “1.” All information in a computer is stored and represented by combinations of “0” and “1.” This information is managed by hexadecimal for convenience. For example, “0110” stands for “16” in decimal notation. Information in a text file in a computer can be observed by a software, such as Binary Editor BZ. Each information is tagged so that the type of information can also be confirmed.

For medical information the DICOM Standard (Digital Imaging and communications in Medicine) is adopted. For example, “08 00 90 10” means the name of manufacturer. However, the number can be described as “00 08 10 90” when little endian is adopted. On the other hand, the order of numbers is not changed in big endian. When an image information file is opened by programing on a computer, the information stored in hexadecimal will be converted into characters and images, which human beings will be able to recognize.

Evaluation of medical images is roughly classified into qualitative and quantitative analysis. Qualitative analysis is detection and differentiation of disease from images, such as X-rays, CT, PET, and MRI images by doctor, especially radiologist, which significantly depends on the doctor’s ability based on his knowledge and experience. Quantitative analysis is analysis by computer processing as to quantitative features, such as signal strength, distribution, shape information, and functional information extracted from these images. This analysis predicts diagnosis and prognosis. Quantitative analysis is expected to realize Computer-Aided Diagnosis (CAD) and personalized medicine.

Quantitative analysis is more objective, but it requires various complicated processes. Quantitative indicators include Standardized Uptake Value (SUV), pixel value statistics in the Region of Interest (ROI), and Apparent Diffusion Coefficient (ADC), which utilizes diffusion phenomenon of water molecules in tissues. Calculation of these indicators requires features extraction, which is made up of data preprocessing, such as segmentation for ROI, features standardization, unification of pixel size by interpolation, features selection, and so on. Furthermore, prediction models must be selected, such as Convolutional Neural Network (CNN). These models predict diagnosis and prognosis from the extracted quantitative indicators. Then, evaluation indicators for prediction by the models will be calculated by the Receiver Operating Characteristic (ROC) analysis. This analysis curve is created from sensitivity, 1-specificity, and cutoff values.

AI can also be applied to pathological diagnosis. AI can judge pathological findings quickly and with the same accuracy as pathologist. However, supervising data need to be carefully created to train AI. For instance, a data image might contain parts that are not suitable as supervising data. One pathological image might include a part that appears to be another disease. Moreover, AI’s judgement can be interfered with by features unrelated to the disease, such as nuclear chromatin.

In addition, there is a problem that the diagnosis reasoning by AI is unknown and becomes a black box. Grad-CAM (Gradient-weighted Class Activation Mapping) can describe the image part that the AI has particularly emphasized to reach the diagnostic imaging, which can teach the judgement reasoning by the AI. Furthermore, it is also considered to add information on clinical findings to the images in order to improve the accuracy. These ideas will make it easier for physicians to use AI for pathological diagnosis and improve the supporting role of AI.

At present, many medical institutions have purchased diagnostic imaging equipment, such as MRI and CT. However, countryside having fewer doctors is extremely short of radiologists who can diagnose these medical images. In the past, radiologists always had to visit rural areas from a big city for diagnostic imaging. The burden on the specialists was heavy, and it took a long time from the examination to the results reporting. At the moment, it is also possible to transmit medical images from rural areas to a specialist in a big city who diagnose them. However, this has not resolved the shortage of radiologists.

Therefore, artificial intelligence (AI) is also expected to play a great role in such remote image diagnosis. AI may be able not only to assist radiologists in diagnosis and report writing but also to distribute medical images to them. However, the present AI does not provide so high the accuracy to detect lesions, and the excessive amount of information can break the Internet. Furthermore, AI tends to provide many difficult images to a small number of particularly talented specialists. It is also necessary to consider the medical fees that medical institutions will receive in case of diagnosis by AI.

Artificial Intelligence (AI) can also perform various tasks in the medical field. For example, AI can carry out classification, which is to learn classification methods from past data to distinguish diseases from unknown data, regression, which is to learn patterns from past data to predict unknown results, clustering, which is to group data based on similarities, and dimension reduction, such as principal component analysis for calculation cost reduction and data visualization. Furthermore, AI can efficiently create images of high quality by denoising.

However, judgement by AI does not guarantee complete accuracy and often leaves its reasoning unclear in the black box. Therefore, application of AI to healthcare involves a lot of problems as to black boxing of clinical AI, safety assurance and performance evaluation, post-marketing performance instability, system error detectability, etc., thus requires strict regulations. In addition, it is necessary for AI manufacturers and physicians to clearly share responsibilities and to jointly carry out risk management to provide patients with medical care of high quality.

The clinical use of medical artificial intelligence (AI) particularly requires the following points: 1, Physician does not need extra work to use the medical AI; Physicians are usually busy with work and they tend to avoid extra time; 2, Medical diagnosis by AI should only point out imaging findings and avoid making a differential diagnosis of the disease; It is not usually possible in clinical practice to make the final diagnosis of the disease based only on information from images; and 3, It is also essential for medical AI to have functions of high accuracy.

In medical AI development, Python is the only programming language, and both Pytorch and TensorFlow (Keras) can be used. However, deep learning by neural network is not the same as human learning, although they are similar. Moreover, for engineers, neural network is by no means a black box, and it is built mathematically. Thus, engineers must understand medical thinking by physicians and design a model that is as identical as possible to thinking by physicians through trial and error from mathematical aspects.

At present, many enterprises are conducting research and development in medical devices that perform diagnostic imaging using artificial intelligence (AI). The examples include automatic setting the shooting position of the human body, identification of a specific anatomic site, and creation of high-quality image with low-dose exposure to radiation by removing noise. If no individual difference is observed as to image contrast or anatomical structure, it is possible to recognize the features in the image according to certain rules. The examples include recognition of the inclination of the midline of the head.

However, machine learning performs image recognition more effectively in many cases because of individual differences in shape. Mainly performed methods are the active shape model and the Adaboost methods. Noise removal in medical images is also made possible by convolutional neural network (CNN). Application of this technology enables conversion into high-quality images from noise-containing images that are taken in a short time and with low-dose radiation exposure. Therefore, the improvement can reduce the examination time as well as the amount of radiation exposure dose during the examination.

Such development of medical devices urgently necessitates reconstruction of regulations on their safety and efficacy as well as privacy protection.

At present artificial intelligence (AI) is widely employed in various industries, such as automobile and translating. Radiomics is also enjoying benefits from the use of AI. Radiomics is to extract a large number of features from radiographic medical images and to select some of them that can greatly contribute to finding disease characteristics and predicting its prognosis and therapeutic response. The process of radiomics includes segmentation for suspicious pathological characteristics. This task can be automated by training machine learning algorithm with datasets. Machine learning is classified into supervised, unsupervised, and reinforcement learning, which are further divided into different kinds of categories.

There are even many types of AI model for carrying out one kind task, such as classification. It is crucial to select the best AI model for each purpose so that AI can acquire the best ability to work for the specific purpose. It is also indispensable to select the most relevant features out of the data as well as hyperparameters that can lead to correct and effective classification. The quality of AI’s classification performance is often measured by sensitivity, specificity and accuracy and visualized by ROC curve.

For example, methylation status and transcriptome subtypes have a great influence for treatment and prognosis of GBM (glioblastoma multiforme). Thus, the classification by machine learning draws great attention for the medical purpose. Here, the XGBoost model was found to be the best classifier and 9 radiomics features were selected for the best performance of the model.

Many issues need to be resolved for successful research and development of artificial intelligence (AI) for diagnostic imaging. At first, medical images are indispensable for production of diagnostic imaging AI. The production needs a certain number of medical image cases that include each disease. There are problems observed here, such as issues as to sharing of images among research institutes and image creation by AI. It is also necessary that the form and the quality of images must satisfy certain uniform requirements. In addition, creating supervising images requires labeling by a specialist.

Medical AI needs to meet the demands of clinical practice. Development of such AI is generally carried out by joint development by clinical specialists and programming experts. For the actual use and commercialization, various issues must be considered, such as the budget, law, intellectual property strategy, including patents, and marketing, not only in the domestic but also in the international market. The reasoning of judgment is particularly important in medical practice. Thus, it is required to resolve the black box problem in AI as much as possible and to ensure its safety and efficacy.