By: Sridhar Nadamuni
Getting radiologists up to speed with artificial intelligence (AI) is essential for successful implementation of new protocols for validation and standardization of AI tools in clinical practice, similar to guidelines for image acquisition used in quantitative analysis. Any clinical use requires stability over a diverse array of settings, equipment, and protocols.
Whether AI is used as an alternative, substitute, or adjunct to radiologic workflows, the application of machine learning, i.e., based on data acquisition and monitoring without prior programming, has been shown to facilitate the detection, segmentation, and classification of images and lesions.1 Patients benefit from automated and rapid detection of critical findings with appropriate imaging quality.
Evidence-based Medicine and AI
In addition to defining the target demographics accurately, all AI solutions must be evidence-based, and all AI tools used in radiology are expected to comply with the STARD (Standards for Reporting Diagnostic Accuracy).2 Similarly, clinical applications based on predictive models require transparency, with follow-up intervals and outcomes delineated in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) statement.3
Clinical Applications—Algorithms are Key
AI algorithms are used to prioritize radiologic screening based on the volume of images and other data. As an accurate, rapid, reproducible, and user-friendly tool, AI may substitute for radiologist in specific cases, for example, estimation of bone age. AI may also be used as an adjunct to radiologist interpretation, based on machine-learning algorithms. For instance, radiologists can perform liver transplantation in patients diagnosed with liver cancer based on the total tumor volume determined via instant lesion segmentation. Radiologists may use AI to distinguish normal from abnormal images. Deep learning algorithms, i.e., based on multiple processing of data with several elements of abstraction, can be used to clearly demarcate abnormal images in a sea of normal findings. Deep learning approaches and convolutional neural networks (CNNs), which are used to analyze visual images, are of prognostic value in radiomic applications.1
Aside from building tools for medical image analysis, it is critical to understand the relevant mathematical calculations, statistical programming, data structures, and appropriate algorithms. Algorithms need to be tailored to practical application. For example, in addition to automated segmentation of liver tumors, the AI tools need to provide information such as tumor type and distribution, and explain the relationship with the blood vessels and staging for clinical application and surgical management.1
AI tools have most recently been used in combination with magnetic resonance imaging (MRI) to detect and segment some of the most common brain cancers including meningiomas 4 and neurodegenerative diseases such as Alzheimer’s disease (AD).
The Role of AI in T1-weighted MRI for Diagnosis of AD
AI classifiers were recently used in combination with MRI to distinguish patients with mild AD from those who were contraindicated for treatment. In brain studies performed using structural T1-weighted MRI combined with an AI classifier based on specific neuropsychological measures, patients with abnormal features were identified almost 2 years prior to a definitive AD diagnosis.5
Indeed, the combination of AI and MRI resulted in greater than 80% accuracy and sensitivity, and 87% specificity. By contrast, MRI data alone yielded no more than 72% accuracy, 69% sensitivity, and 75% specificity. Abnormal features detected in the temporal and medial-temporal cortex may represent biomarkers of clinical interest. Abnormalities detected 24 months before a definitive diagnosis of AD may provide opportunities for appropriate intervention.5
Automated, Sensitive, Objective and Cost-effective Intervention
MRI is considered less expensive than positron emission tomography (PET), and is a non-invasive and frequently used diagnostic modality worldwide. MRI is recommended for the identification of neuronal degeneration in AD and to monitor AD progression in clinical trials.6
However, the signs of mild neurocognitive degeneration are not always perceptible to radiologists. Further, it is not always possible to predict AD development based on MRI alone.
AI technology represents a promising and effective tool for automated, objective and highly sensitive assessment of imaging studies. Techniques based on machine learning and pattern recognition are of neuroimaging interest in unraveling novel data.7
AI algorithms extract information without prior knowledge of the location of coded data in the images. In addition, they amalgamate the information found within and across several different variables, which can be used to develop mathematical models for automated prediction of a patient’s diagnosis. Such tools are of immense value in early diagnosis, when the tell-tale signs of disease are not obvious.8
Several MRI studies of the brain using machine learning reported automated diagnosis and prognosis of early-stage AD.9 AD has been diagnosed in more than 60 million people worldwide, and is estimated to affect 75.6 million in 2030, and nearly 135.5 million by 2050.10,11
A recent Harvard/MIT study demonstrated the ability of machine learning to enhance MRI scans.The AUTOMAP (automated transform by manifold approximation) obtained using 50,000 MRI brain scans from the Human Connectome project increased image reconstruction with less signal-to-noise ratio than the conventional MRI, obviating the need for repeated patient visits, and facilitated accurate clinical decision-making.12
Recent FDA Approvals
The United States Food and Drug Administration (FDA) recently cleared several image analysis applications using AI.13
The Arterys Oncology AI suite represents the first FDA clearance for oncology imaging via deep learning, for radiologic measurement and tracking of potential cancers. Initial deep learning approaches involving Arterys were designed for liver MRI and computed tomography (CT) scans as well as for lung CT scans. Using a simple web browser, clinicians can now diagnose and quantify the presence or absence of lung nodules and liver lesions along with their key characteristics. Deep learning technology enables instant segmentation of lung nodules and liver lesions. The accuracy is comparable to manual evaluation by clinicians, and the segmentations can be edited subsequently as needed.14
1. Canadian Association of Radiologists White Paper on Artificial Intelligence in Radiology. Canadian Association of Radiologists Journal. https://doi.org/10.1016/j.carj.2018.02.002. Accessed July 20, 2018.
2. STARD 2015: An Updated List of Essential Items for Reporting Diagnostic Accuracy Studies.Radiology. https://pubs.rsna.org/doi/10.1148/radiol.2015151516. Accessed July 21, 2018.
3. Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD): The TRIPOD Statement. Annals of Internal Medicine.http://annals.org/aim/fullarticle/2088549/transparent-reporting-multivariable-prediction-model-individual-prognosis-diagnosis-tripod-tripod. Accessed July 21, 2018.
4. Fully automated detection and segmentation of meningiomas using deep learning on routine multiparametric MRI.European Radiology. https://doi.org/10.1007/s00330-018-5595-8. Accessed July 21, 2018.
5. MRI Characterizes the Progressive Course of AD and Predicts Conversion to Alzheimer's Dementia 24 Months Before Probable Diagnosis. Frontiers of Aging in Neuroscience 2018 May 24;10:135. doi: 10.3389/fnagi.2018.00135. eCollection 2018.
6. Amyloid-related imaging abnormalities in amyloid-modifying therapeutic trials: recommendations from the Alzheimer's Association Research Roundtable Workgroup. Alzheimers & Dementia. https://www.alzheimersanddementia.com/article/S1552-5260(11)02506-4/fulltext. DOI: https://doi.org/10.1016/j.jalz.2011.05.2351. Accessed July 21, 2018.
7. Machine Learning in Medical Imaging. IEEE Signal Processing Magazine. https://ieeexplore.ieee.org/document/5484160/?reload=true. Accessed July 21, 2018.
8. Frontiers for the Early Diagnosis of AD by Means of MRI Brain Imaging and Support Vector Machines. Current Alzheimer Research. https://www.ncbi.nlm.nih.gov/pubmed/26567735. Accessed July 21, 2018.
9. Combining multiple approaches for the early diagnosis of Alzheimer’s disease. Pattern Recognition Letters. DOI10.1016/j.patrec.2016.10.010. Accessed July 21, 2018.
10. Dementia Statistics (2015). Alzheimer’s Disease International. http://www.alz.co.uk/research/statistics. Accessed July 21, 2018.
11. The Global Impact of Dementia, An Analysis of Prevalence, Incidence, Cost and Trends. World Alzheimer Report (2015). http://www.alz.co.uk/research/WorldAlzheimerReport2015-sheet.pdf. Accessed July 21, 2018.
12.Image reconstruction by domain-transform manifold learning. Nature. doi: 10.1038/nature25988. Accessed July 21, 2018.
13. Artificial intelligence in radiology: Hype or hope? Applied Radiology. https://www.appliedradiology.com/articles/artificial-intelligence-in-radiology-hype-or-hope. Accessed July 20, 2018.
14. Arterys Receives First FDA Clearance for Broad Oncology Imaging Suite with Deep Learning. https://www.prnewswire.com/news-releases/arterys-receives-first-fda-clearance-for-broad-oncology-imaging-suite-with-deep-learning-300599275.html. Accessed July 20, 2018.
15. Artificial Intelligence helps early detection of dementia. EUREKA EUROSTARS PROJECT 8234, BrainIQ. https://www.eurostars-eureka.eu/content/artificial-intelligence-helps-early-detection-dementia. Accessed July 20, 2018.
16. A Survey on Deep Learning in Medical Image Analysis. https://arxiv.org/abs/1702.05747. Accessed July 22, 2018.