Artificial Intelligence in Medical Images – Opportunities and Challenges
by Yi Pan, Regents’ Professor and Chair of Computer Science Department at Georgia State University, Associate Editor of Journal of Computer Science and Technology
Artificial Intelligence (AI) is the science of mimicking human intelligences and behaviors. Machine Learning (ML), a subset of AI, trains a machine how to use algorithms or statistics to find hidden insights and learn automatically from data. Deep learning (DL) is one of machine learning methods where we use deep neural networks with advanced algorithms such as auto-encoding or convolution to recognize patterns in data. Deep learning has become very successful recently due to the availability of huge data and powerful supercomputers. Many applications such as speech and face recognition, image classification and natural language processing suddenly took great leaps due to the advance of deep learning.
Medical images have been used by doctors for many decades to diagnose diseases. Due to the capabilities of DL, using DL to process medical images and recognize patterns in it is a natural application and also an economical way of handling these images due to the higher labor cost of medical doctors and cheaper computing power. There are three types of medical images. Radiology images consists of MRI images, X-Ray images, PET, and CT images and are at the organ level. Radiologists read these images on a film through finding particular features such as black spots to diagnose a disease. Histopathology images are obtained through a microscope to capture the entire slide with a scanner and to save it as a digital image which contains details at the tissue and cell level. A pathologist observes the stained specimen on the slide glass using a microscope to perform a diagnosis. Molecular images are obtained via electron microscopes and able to display details at a molecular level such as light atoms in the vicinity of heavy atoms or measurable bond lengths inside a cell. Molecular imaging differs from traditional imaging in that biomarkers can display particular disease targets or pathways. Histopathology images and molecular images are obtained invasively, while radiology images are noninvasively. While ML algorithms in MRI, X-Ray, PET, and CT images are used to complement the opinion of a radiologist, a computer algorithm for histopathology images are developed for disease detection, diagnosis, and prognosis prediction to complement to the opinion of the pathologist. Molecular images can also help biologists to identify disease biomarkers and detect diseases such as cancer. Besides improving diagnosis of diseases, AI technologies also contributes to improving the treatment of these diseases through optimizing the pre-clinical and clinical tests of new medication. The progression prediction and earlier and more precise diagnosis means saving more healthcare costs and lives. Hence the AI technology has a major economic and societal impact and received a lot of attention recently.
As a human being, a doctor learns from textbooks or his mentor on what features are important to identify a disease. If his knowledge is not complete or the knowledge from the textbooks is not perfect, he makes errors. On the other hand, ML can learn the unknown knowledge and discover new important features automatically through training. Features and classifiers are simultaneously optimized in deep learning thus producing outstanding results. Besides feature extraction and classification, ML algorithms can also perform denoising, segmentation, registration, clustering, and diagnosis. A computer algorithm such as a neural network, once trained, can perform diagnosis quickly and economically. If the AI system can also perform a better job in prediction, its adoption in future diagnosis is almost a certainty.
However, many obstacles still need to be overcome. Besides common problems in AI such as high training time, big memory requirement, and privacy and security issues, they have several unique issues. 1. AI requires a lot of training data to achieve a good prediction model. But medical image data are hard to get and people are not willing to share due to their high production costs. Maybe it is time for medical image data to become an intellectual property and commodity. 2. Due to different requirements, each domain requires a different model and there is no universal theory to guide model and parameter selection. It is still a trial and error science, like “alchemy” lacking science and theory. 3. Multi-modality learning has been proposed to be a better model than single modality since different data sources have different strengths and the final results after fusion are usually better. For example, to diagnose a particular disease, we can use MRI, functional MRI, CT, histopathology images, and molecular images plus genomic data, microarray data, and hand-crafted features. When these prediction results are ensembled to achieve the best result, we require the data are from the same group of patients. This requires we test these patients using different methods at the same time, a daunting task to do for any hospital. 4. Current AI boxes like DL are black boxes; you get the prediction results, but you do not know how they achieve the results. Understanding the prediction process and extracting rules of the results are the next important topic. Is the prediction result due to a black area or a smooth circle on the X-ray image? Explainable AI is especially important for medical images. 5. How to exploit human experience within the AI system is tricky. How can a doctor inject his prior knowledge into a DL system to improve the prediction result? 6. Finally, there is a legal issue in using medical images to predict a disease. Like human being, AI is bound to make a mistake that harms a patient at some point. Since the law is not mature yet, a wrong decision from AI would potentially cause a lawsuit. It is unlikely in the near future that a commercial AI algorithm would be deployed by a hospital without the blessing of the government. Right now, AI systems are still used as an aid. When AI produces a much higher diagnosis precision, with the protection of related law, the dream that AI systems could be used in medical care may become true. These challenges create a huge number of opportunities for people in both computer science and health care.
The book
About the author
Dr. Yi Pan is currently a Regents’ Professor and has served as Chair of Computer Science Department at Georgia State University since January 2006. He has also served as an Interim Associate Dean and Chair of Biology Department during 2013-2017. Dr. Pan joined Georgia State University in 2000, was promoted to full professor in 2004, named a Distinguished University Professor in 2013 and designated a Regents' Professor (the highest recognition given to a faculty member by the University System of Georgia) in 2015.
Dr. Pan received his B.Eng. and M.Eng. degrees in computer engineering from Tsinghua University, China, in 1982 and 1984, respectively, and his Ph.D. degree in computer science from the University of Pittsburgh, USA, in 1991. His profile has been featured as a distinguished alumnus in both Tsinghua Alumni Newsletter and University of Pittsburgh CS Alumni Newsletter. Dr. Pan's current research interests mainly include bioinformatics and health informatics using big data analytics, cloud computing, and machine learning technologies. Dr. Pan has published more than 450 papers including over 250 journal papers with more than 100 papers published in IEEE/ACM Transactions/Journals. In addition, he has edited/authored 43 books. His work has been cited more than 13,000 times based on Google Scholar and his current h-index is 58. Dr. Pan has served as an editor-in-chief or editorial board member for 20 journals including 7 IEEE Transactions. Currently, he is serving as an Associate Editor-in-Chief of IEEE/ACM Transactions on Computational Biology and Bioinformatics. He is the recipient of many awards including one IEEE Transactions Best Paper Award, five IEEE and other international conference or journal Best Paper Awards, 4 IBM Faculty Awards, 2 JSPS Senior Invitation Fellowships, IEEE BIBE Outstanding Achievement Award, IEEE Outstanding Leadership Award, NSF Research Opportunity Award, and AFOSR Summer Faculty Research Fellowship. He has organized numerous international conferences and delivered keynote speeches at over 60 international conferences around the world.
Line