Machine learning broadly classified into supervised learning, unsupervised learning, and reinforcement learning. Supervised learning is mainly used for regression and classification, and unsupervised learning is mainly used for grouping and summarizing information, respectively. Reinforcement learning is utilized for games such as shogi and go as well as robots to learn walking and is performed without using statistics. Supervised learning needs the use of data that include correct answers. Reinforcement learning includes reception of rewards or estimations for trials. There is also semi-supervised learning, which is a combination of supervised and unsupervised learning.
Typical analysis methods for supervised learning are regression analysis, nearest neighbor method, decision tree, and support vector machine, and typical analysis methods for unsupervised learning are k-means clustering, association analysis, and dimensionality reduction. At present, deep learning is often used. It uses multi-layer perceptron and backpropagation to learn the characteristics of the data by itself. Activation functions have been improved, which can prevent gradient disappearance at the time of backpropagation.
Regression analysis is a typical example of supervised learning and consists of processes to create a scatter plot, to estimate the model, to evaluate efficacy of the model, and to evaluate correlation among independent variables. Evaluation of efficacy is performed by calculation of coefficient of determination and multiple correlation coefficient, statistical testing, and residual analysis. Multicollinearity means strong correlation between independent variables. Regression analysis is interpretation-oriented and gives rise to a simple model. On the other hand, deep learning is accuracy-oriented and requires complicated network as a black box. As a result, it is extremely difficult for humans to interpret.
Dimension compression belongs to unsupervised learning and uses principal component analysis to group explanatory variables into the principal component having the highest contribution ratio and the other components. In case two components are made, the eigenequation is solved to lead to the first component of the larger eigenvalue and the second component of the smaller eigenvalue.
Clustering is also classified as unsupervised learning and a technique to divide a large amount of unlabeled data into several sets of data according to their similarity. Clustering is classified into hierarchical clustering and non-hierarchical clustering. A typical example of the former is dendrogram and a typical example of the latter is the K-means method.
To date, artificial intelligence (AI) is also used for research conducted in the field of nuclear medicine, and some of them have been published in leading academic journals. For example, AI is used for automatic detection, contour determination, and qualitative diagnosis for diseased lesions and organs. Furthermore, it is also researched to find the best measures to deal with aging by organ by predicting the age from normal images, although many normal images are disposed of without attention in the medical practice. In addition, research is also performed to find a method to reduce the radiation exposure on patients for diagnostic imaging by producing high-quality images from medical images containing noise taken at low-dose exposure. This method has a greatly advantageous feature that supervising data can be created without the need of labeling by a specialist because a low-quality image can be produced by adding noise to a high-quality image that already exists. Although the usefulness of AI has been proven to some degree, AI can be hardly used in clinical practice as long as the judgment process of AI remains a black box. It is hoped from now that explainable AI will be developed.
In the field of nuclear medicine, artificial intelligence (AI) particularly plays a great role in diagnostic imaging support, image generation, and noise reduction. It is expected that double check using high-precision diagnostic imaging by AI will reduce workload of physicians as well as contribute to early detection of illness. At present, software for diagnostic imaging have been developed. Here, there is an important step of data pre-processing, which requires complicated processes, consisting of standardization, smoothing (noise removal), image-to-image difference correction, mask processing, and count standardization. Thereafter, they are followed by statistical analysis, etc.
Image generation is to generate one image from another of different kind. For instance, a PET image of a patient is used to predict and produce a CT image. As a result, a patient will have to take a smaller number of tests. Research is also being conducted to create a high-quality image from a low-quality image that was taken with low-dose radiation exposure by noise removal. This technology will be able to reduce radiation exposure suffered by a patient during examinations.
A method is being developed to apply artificial intelligence (AI) to bone scintigraphy to automatically calculate the BSI (Bone Scan Index) value. The BSI value is a quantification value as to the state of cancer metastasis and has a significant relation with the prognosis of patients. Different BSI values are calculated for different combinations of cancer metastases and treatments. Thus, they can predict the effect of treatments on cancer metastasis.
However, it must be noted that calculation of the BSI value is affected by the testing equipment, the dose of radiopharmaceuticals, the patient’s conditions other than ones as to the disease, such as urine bag, and image noise, and so forth. Moreover, it is strongly recommended to analyze using the same software for following up on the same patient.
The Analysis flow is composed of performance of bone scintigraphy, classification of the skeletons in the image according to the Morpho theory, detection of high integration parts, normalization of the data, quantification of the high integration parts, and classification of the quantified data and calculation of the BSI value by AI. Research is also conducted to add other images, such as CI and blood data for this analysis.
Artificial intelligence (AI), especially Convolutional Neural Network, can also be used for medical images to carry out image classification, object detection, segmentation, super-resolution, denoising, and similar image generation. Contrary to object detection, segmentation detects objects having irregular shapes. Segmentation has undergone evolution from semantic and instance into panoptic segmentation, which enables labeling every pixel and recognizing objects individually. Each evaluation method is also different.
However, it is time-consuming to create supervising data in segmentation. Object detection only requires rough enclosure of the object in a bounding box, whereas segmentation requires accurate definition of the object to avoid wrong learning by AI. Moreover, medical images are saved in a special image format called DICOM and the size is also various. Thus, the image format and size must be converted for the purpose of AI learning. Usually, data augmentation is also needed. After that, the hyperparameters (learning parameters) are fixed, and the AI model will be trained and evaluated.
Research in image classification by artificial intelligence (AI) firstly requires determination of the object of classification, the purpose of classification, and the image data to be used based on past literature. Higher Accuracy of AI will heighten the difficulty as well as the significance of the research. Generally, a research topic is selected from the strong fields of the research institution or hospital that the researcher belongs to.
The next step requires determination of the type, cross-sectional direction, range, size, etc. of the image data to be used, which will further undergo data augmentation according to the required number of data images. Then, the image data will be split into training data and test data.
In many cases, convolutional neural network is employed for transfer learning or fine tuning of an AI model. Then, the optimizer (often Adam) and the hyperparameters (mini-batch number, epoch number, and learning rate) are empirically determined for the model training. Finally, the performance of the trained model will be evaluated with test data.
Research using AI is often conducted based on empirical rules, thus requires repeating trial and error for understanding of the application scope and limitation of AI.
The role of video processing by artificial intelligence (AI) comprises post-hoc analysis of video recording with regard to way of walking and motion forms, real-time analysis as monitoring, abnormality detection, and sensor, and understanding of the meaning of videos, such as video summaries. A moving image is composed of continuous static images, and the number of images per second is described by fps (frame par second). Each image is disposed of by image classification, object detection, and segmentation, and they will be joined together into one series so that AI appears to be analyzing one movie video.
Video is also used in healthcare, and the use of AI is also expected. For example, AI can predict the correct positional relationship between an X-ray device and a subject. That is, AI will assist in X-ray photography. However, learning video requires production of supervising image for each image. Furthermore, it is necessary to plan how to process the obtained data in order that AI will be able to carry out the necessary learning.
Natural Language Processing (NLP) is solving problems related with natural language using computer. NLP consists of morphological analysis used for machine translation and document summarization, construction analysis applied to document classification and sentiment analysis, semantic analysis used for chatbots and native expression depiction, and context analysis applied to document generation and relationship extraction. Morphological analysis is made up of the process of word segmentation, stemming or lemmatization, and part-of-speech tagging. Rare words are tokenized as subwords. Special care needs to be taken for unique medical terms.
At present, character information is quantified, which is called distributed representation. Such vectorization of characters is performed by Word2vec, fastText, Doc2vec, and the like. Recently, deep learning is also used in the field of NLP. Initially, recurrent neural networks (RNNs) were used, but they have the vanishing gradient problem. Today, LSTM (Long short-term memory), Seq2seq (Encoder-Decoder model) with Attention, Transformer, and BERT (Bidirectional Encoder Representations from Transformers) have been developed, which has enabled advanced NLP. BERT achieves high general-purpose by performing pre-training with unlabeled data and fine-tuning with a small amount of labeled data.
At present, artificial intelligence (AI), particularly deep learning, is also being used for drug discovery. In bioinformatics AI predicts the structures and interactions of the protein. In cheminformatics AI performs automatic designing of candidate components and prediction of their pharmacokinetics and toxicity. In the new drug development from existing drugs, the new ways to use the existing drugs can be discovered by network-based algorithms, gene expression-based algorithms, and docking simulation-based drug design, and so forth. Moreover, AI can predict the protein three-dimensional structure and propose new drug candidates having a three-dimensional structure that easily binds to the protein’s structure. In addition, AI can predict adverse effects of a drug and classification that cannot be expected by humans.
On the other hand, AI drug discovery also has many problems to be resolved, including the safety, the accountability and responsibility for erroneous predictions, protection of personal information, the time for data collection, the accuracy of supervising data, and intellectual property rights of enterprises. On top of that, medical data have been collected primarily for the purpose of clinical treatment. Therefore, it is disputed how such medical data should be used for AI learning.
Application of artificial intelligence (AI) in the healthcare field involves various challenges to be resolved at each AI development stage. At the stage of data exploration and analysis, poor annotation and labeling, class imbalance, and bias are often observed in medical datasets. Appropriate data collection and preparation is indispensable. For building AI model, overfitting must be avoided, and robustness and generalizability must be secured. Selection of hyper parameters is the key to deal with these problems. At performance check, the reliability of the performance by AI model must be periodically assessed.
At present, AI is widely being developed in radiology, particularly for diagnostic imaging. The use of AI for diagnostic imaging can be roughly classified into three purposes: detection, segmentation, and classification. Detection is conducted to identify target lesions of interest in medical image. Segmentation is performed to describe the boundaries of target lesions of interest. Classification is carried out to categorize target lesions. Different AI model will be selected for each purpose, and various Python packages for machine learning can be also used for AI development in the healthcare field.
At present artificial intelligence (AI) is used in various fields, and particularly, it has higher ability than human beings in image recognition. Images of poor quality includes noise and are often difficult for humans to look at. AI can reproduce a clear image by removing this noise. Moreover, AI has ability to recreate an object that is difficult to observe for human eyes because the image is taken in a dark spot, or that is hidden by an obstacle.
Therefore, such image recognition by AI is applied to different industrial fields, including healthcare, transportation, business, and environment. For example, in subway routes in the civil engineering field, it is always necessary to check damaged portions in the tunnel for its safety. The distance of the subway passage is extremely long, and it is considerably difficult for humans to check every place directly. In this case, AI can enable us to find places effectively that need repair by recognizing the damaged parts and visualizing them with coloring.
Recently, artificial intelligence (AI) has been rapidly developed and has begun to be used in various fields. However, AI must mainly overcome three problems in order to become widely practicable: 1, The accuracy of AI should be equivalent to that of specialists in this art; 2, The process of judgements by AI should be accountable; and 3, AI model should be trainable by a small amount of data. Particularly, it is extremely difficult from the point of privacy protection to collect patients’ data in the healthcare field.
In recent years, Generative Adversarial Networks (GAN) has been invented and enabled to create a large number of similar images out of single image data. That is, it has become possible to generate a large amount of similar image data from a small amount of image data to produce an AI model having high accuracy. Furthermore, technology has been developed to sense human cognitive process, and research is being conducted to enable AI to follow and visualize this process. This will make it possible to understand the process by which AI makes decisions.
Medical device is a device used for diagnosis, treatment, and prevention of diseases, or one intended to affect the function of the human body. Medical device program is a program that has the purpose of a medical device and may have adverse effects on the life and health of the patients and users if it fails to function as intended (SaMD: Software as a Medical Device). It includes clinical artificial intelligence (AI).
Theoretically, clinical AI products can be developed according to the type of disease and the type of healthcare service, such as interview, examination, diagnosis, treatment, drug discovery, informal care. However, the reality is that the market only accepts AI products that satisfy the needs of actual clinical practice and legal regulations.
It takes a huge amount of development cost and a certain period of time to obtain manufacturing and marketing approval as well as to ensure the quality and safety in order that an AI product can be put on the market as a medical device. The advantages are that the product can be used clinically and that a barrier to entry for the product can be easily built.
Recently, many applications related to medical imaging in artificial intelligence (AI) have been researched and developed. Three main reasons are observed: 1, innovation of technology, such as Convolutional Neural Network in image recognition that is superior to the human capacity; 2, improvement of the processing capability of computer as well as storage; and 3, uniformed digitalization of clinical imaging. The application to medical imaging is mainly classified into three categories: 1, assisted interpretation; 2, additional insights; and 3, augmented image.
Assisted interpretation assists for diagnostic imaging, such as automatic anomaly detection. Additional insights discover a new association between image findings and clinical outcomes. Augmented image improves the quality of images, including noise removal. However, each application also has problems that must be resolved. Assisted interpretation requires production of supervising data by specialist in diagnostic imaging as well as appropriate transfer learning. Additional insights require accurate reflection of the characteristics of the tissues from the numerical data that were analyzed AI. Augmented image requires not only sufficient training of the AI model, but also sufficient image information held in test data.