E-Bulletin, Issue 3 from June 2, 2020
RSNA Goes Virtual; Artificial Intelligence Provides Differential Diagnoses Almost at Neuroradiologists’ Level
Welcome to our third monthly newsletter on neuroradiology and AI. We hope you find it helpful. Please feel free to share it with others.
If you would like to be informed when the new issue is out, please sign up here.
Virtual RSNA: RSNA 2020 will move to a virtual format this year
Due to the coronavirus pandemic, RSNA 2020: Human Insight/Visionary Medicine will be held online instead of in Chicago. Already, more than 11,000 scientific and educational abstracts were submitted. The event was expected to draw 50,000 attendees from 137 countries, as the world’s largest radiology conference. The meeting will be from 29 November through 5 December. Registration for the virtual conference opens on 22 July. Find exhibitor information here.
Stroke clinical trials: 5 trials presented at ESO-WSO 2020
Despite the ESOC-WSO conference being postponed to later this year, the organizations have opted to continue keeping stroke communities updated on recent achievements in stroke analysis and treatment. Five clinical trial results were presented during the ESO-WSO 2020 webinar focusing on advances in stroke care. Presentations included results on major clinical trials assessing the efficacy of fluoxetine; a comparison of direct endovascular thrombectomy and thrombectomy (with or without intravenous alteplase); a trial of endovascular therapy in basilar artery occlusion; and a trial comparing the effect of atorvastatin or placebo, and amlodipine-based or beta-blocker-based regime on the 20-year incidence of stroke or dementia. You can watch the webinar here.
Neuroradiology and AI in the news
Machine-learning algorithm on functional MRI (fMRI) predicts whether patients in cardiac arrest-induced comas will regain consciousness. Patients who experience out-of-hospital cardiac arrests and admitted to the intensive care unit, often have a post-anoxic brain injury. This leads doctors to withdraw care. Those in a coma may undergo clinical examination, structural and diffusion neuroimaging, biomarker identification and electrophysiologic testing. Researchers said this is accurate in predicting poor outcome, but these methods are not sensitive enough to predict consciousness recovery. The study used machine learning classification methods, finding that resting-state fMRI led to high accuracy in identifying patients with positive and negative prognosis, in those with an indeterminate prognosis after standard multimodal testing. Their study, published in the American Journal of Neuroradiology, included 17 patients, 9 of whom regained consciousness while 8 remained comatose.
Predicting Alzheimer’s Risk with AI, MRI and patient data. There are no foolproof ways to definitively diagnose a person with Alzheimer’s disease without examining pathology. Current methods combine patient history, neuropsychological testing and MRI studies, but researchers find that application is varied and lacks sensitivity and specificity. This study in Brain reports results from an interpretable deep learning strategy, combining unique Alzheimer’s disease signatures from MRI multimodal inputs, gender, age, and Mini-Mental State Examination scores. Their model trained on cognitively normal and clinically diagnosed Alzheimer’s disease subjects, using the Alzheimer’s Disease Neuroimaging Initiative dataset with 417 records, validating it on three independent cohorts. Researchers found that the framework provided a clinically adaptable strategy for incorporating available imaging such as MRI, to develop neuroimaging signatures for the disease and diagnosis, including those with disease risk en route to a diagnosis.
Machine learning used for stability assessment of intracranial aneurysms. Research published in Translational Stroke Research tested the hypothesis that machine learning could be used to help assess stability of unruptured intracranial aneurysms (IAs). They developed multiple machine learning models, and assessed 1,539 patients with stable IAs, and 528 with unstable IAs. The models incorporated patient-specific clinical features and aneurysm morphological features. They generated support vector machine (SVM), random forest (RF), and feed-forward artificial neural network (ANN) models, comparing the discriminatory performances of these models to the statistical logistic regression (LR) model and the PHASES score for IA stability assessment. They found that all machine learning models performed in a superior fashion. Researchers concluded that machine learning has the potential to augment the clinical decision-making process for IA stability assessment.
Artificial intelligence is close to neuroradiologists’ accuracy in differential diagnoses on brain MRI. Research published in Radiology focused on using AI to create differential diagnoses for common and rare diseases using brain MRI, comparing the results to those of radiologists with varied experience levels. Researchers used 19 common and rare diagnoses from a 10-year period ending in 2018. The AI system combined domain-expertise methodologies (deep learning and Bayesian) with data-driven ones. They compared AI accuracy to results from radiology residents, general radiologists, and neuroradiology fellows and academic neuroradiologists. They found that AI system performance was not affected by disease prevalence. The AI system used on brain MRI was close to the top three differential diagnoses in accuracy, compared to neuroradiologists, and AI had a higher accuracy compared to less specialized radiologists.
AI algorithms continue to gain scientific validation, and can be helpful in supplementing the work of all radiology experience levels. AI can provide high quality data, alerting physicians to prioritize certain patients in the workflow. Increasingly, it’s important to understand AI capabilities and begin incorporating these tools in the healthcare environment.