Artificial Intelligence Based New Era of Optical Molecular Imaging
by Jie Tian, dean of Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, Beihang University, Editor in Chief of Visual Computing for Industry, Biomedicine, and Art
Artificial intelligence (AI) is a science that endows artificial creations with cognitive, thinking and other behavioral patterns through human intervention. Through computer science, AI relies on bio-coding or physical coding and attempts to build bio-intelligence. Machine learning (ML) technology, which is an important part of AI, uses an artificial pattern recognition system or data-driven-based presentation learning method to achieve intelligent behavior. In particular, the presentation learning method, which is based on implicit features learned from raw data, improves the complex data analysis performance of an AI system. Although this method has a powerful semantic representation capability, it relies on large amounts of data. However, with the development of computer storage technology, communication technology, and computer computing performance, it is now possible to collect and manage large-scale data that allows the bottleneck of the presentation learning method to be solved. Thus, presentation learning, especially the deep learning (DL) method, is now widely used. The DL-based method uses neural networks to extract features in different scales and spaces from natural data and then expresses the features in a cascading way. Because the DL performs well on high-dimensional data, it is being used in many domains, including image recognition, speech recognition, and natural language processing. It is also used in network information security, industrial data processing, and drug molecular research. Now we will introduce the application of AI in optical molecular imaging (OMI).
OMI is a molecular imaging technology that uses an optical signal as the imaging medium. Based on the biochemical characteristics of tissue in organisms and related reactions, OMI acquires imaging signals by observing the distribution of photons on the surface of organisms or the interference of the laser with biochemical reactions. Thus, it performs well for detecting biomedical activities. With the advancements in computer science, molecular biology, and bio-photonics, new methods, such as the molecular probe technology, optical signal acquisition, and amplification technology, have been gradually applied in NIS, significantly improving the near-infrared light acquisition capability of OMI. These improvements allow OMI to observe pathological information in organisms through optical signals.
Due to OMI's need for image data processing and analysis, AI has been applied to different optical modalities. It provides a new image processing tool for disease classification, lesion detection, segmentation, three-dimensional (3D) visualization, and tomographic image reconstruction. For example, in the field of image enhancement, Daniele et al. proposed a novel synthetic data generation approach to construct the ground-truth data and used these data to evaluate the performance of different exemplar-based deep neural networks in obtaining optical images with super-resolution. Recently, Chong et al. developed a postprocessing method for fluorescence image enhancement, which employs a generative adversarial network to improve the image resolution. To overcome the drawback of fake texture generation in traditional neural networks, they proposed a total gradient loss for network training and applied a fine-tuning training procedure to further improve the network architecture. For image registration, Mountney et al. proposed a context-specific-based feature descriptor and used a decision tree to present the feature point and match these points based on their likelihood. They used this method to describe the 3D space of endoscopic images in MIS and reported that the registration results are robust to drift, occlusion, and changes in orientation and scale.
AI can still be utilized in disease classification based on medical images. For example, Guillaume Lemaĭtre et al. adopted a combination of random forest with local binary pattern features and different mapping strategies to classify Diabetic Macular Edema versus normal subjects. Manually defined image features can effectively characterize certain characteristics of a disease, thereby improving the recognition of linear and nonlinear classifiers. However, the pathological significance of image features needs to be further explained to demonstrate the reliability of the classification results. Lee et al. used VGG-16 to automatically classify AMD data. They reported that their network, which was trained with 100,000 OCT B-scan images, achieved an area under curve (AUC) of 0.97 in the validation.
Although many researches have tried to apply the ML technology in OMI and achieved promising breakthroughs, the application of ML-based methods in clinical surgery still requires more theoretical research and clinical experiments. Because the structured data of most clinical applications are insufficient to support the complexity of data-driven-based ML technology, further research is needed for collecting more data and designing novel ML methods based on small-scale data learning. Furthermore, there is still no ideal method that explains the mechanism of a neural network. The features extracted by exciting methods are difficult to illustrate the relevant theories of medicine or optical imaging. This is an urgent problem of neural network that needs to be overcome.
However, ML-based AI still has a very broad space of research and applications. With the expansion of standardized data, the generalization and robustness of ML methods, such as DL, are expected to improve. Furthermore, attention mechanism and other strategies have been proposed to visualize the prediction basis of neural networks, which provide a feasible scheme for network interpretation. With the continuous developments in related research, the application of ML in OMI remains a promising technique.
The book
About the author
Dr. Jie Tian has been elected as the Fellow of ISMRM, AIMBE, IAMBE, IEEE, OSA, SPIE, and IAPR. He is dean of Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, Beihang University. He serves as the editorial board member of European Radiology, IEEE Transactions on Medical Imaging, IEEE Transactions on Biomedical Engineering, IEEE Journal of Biomedical and Health Informatics, and Photoacoustics, and he is the Editor in Chief of Visual Computing for Industry, Biomedicine, and Art.
Dr. Tian has more than 100 granted patents in China and six patents in the United States. He is the author of over 300 peer-reviewed journal articles, including publication in Nature Biomedical Engineering, Nature Communications, Science Advances, Advanced Materials, Journal of Clinical Oncology, Gut, Gastroenterology, Clinical Cancer Research, Radiology, and many other journals. His publications received over 26,000 Google Scholar citations (H-index 79). He is the editor of five academic books in the field of medical imaging, and his book, Molecular Imaging Fundamentals and Applications, is sold online for over 58,000 times. He is one of the founders of Chinese Society for Molecular Imaging (CSMI), and was elected as the first president of CSMI (2010-2017). He received numerous awards, including five national top awards for his outstanding work in medical imaging and biometrics recognition.
Line