The growth of electricity starting about a hundred years ago transformed every major industry, ranging from healthcare, to transportation, to manufacturing, to agriculture. AI is similarly poised to do the same. In this presentation, you’ll hear about recent trends in AI. You’ll also see several case studies in applications of AI to healthcare data, on problems ranging from reading EKG to prioritizing palliative care to reading chest x-rays. This presentation will also discuss general AI principles underlying these applications, and discuss future opportunities. 

This presentation includes work jointly done with Stanford Medical school faculty Sanjay Basu, Stephanie Harman, Curt Langlotz, Matt Lungren, Nigam Shah, and Mintu Turakhia, as well as with Anand Avati, Awni Hannun and Pranav Rajpurkar. 

Biobank resources like the UK Biobank contain a wealth of phenotypic information for each subject compared to traditional ascertainment-based studies. However, specific diseases are generally represented near their population prevalence in biobanks which can hamper studies of individual diseases. New statistical methods that can jointly analyze multiple phenotypes are therefore needed to fully take advantage of biobank resources with linked health records and other phenotypic measurements. We have developed a suite of methods to study disease genetics in the context of population-scale biobanks that jointly leverage information from multiple phenotypes to estimate genetic parameters such as genetic correlations, model disease risk, and identify likely disease-associated variants and genes. These methods use GWAS summary statistics as input which allows easy meta-analyses across biobanks and with previous studies while protecting the privacy of participants. This presentation will describe these novel approaches and their application to data from the UK Biobank.

Vast amounts of molecular data characterizing the genome, epi-genome and transcriptome are becoming available for a wide range of cancers. In addition, new computational tools for quantitatively analyzing medical images and are creating new types of phenotypic data.  Now we have the opportunity to integrate the data at molecular, cellular and tissue scale to create a more comprehensive view of key biological processes underlying cancer. Moreover, this integration can have profound contributions toward predicting diagnosis and treatment. I will discuss our current work developing methods for multi-omics data modeling and  show examples for non-small cell lung cancer and glioblastoma.

Radiological diagnosis and interpretation is ready for an overhaul. Radiologists spend countless hours on tasks that are onerous and error-prone, resulting in high costs and frequent misdiagnoses. Arterys is working to address these deficiencies, using deep learning to vastly improve the speed and consistency with which radiologists read cardiac MRI studies. Our first product, Arterys Cardio DL, is the first technology ever to be cleared by the FDA that leverages cloud computing and deep learning in a clinical setting. We discuss the technology behind the software and how we proved its safety and efficacy to secure FDA clearance in the United States and the CE Mark in Europe.

Visualization and exploration of high-dimensional data is a ubiquitous challenge across disciplines. Widely-used techniques such as principal component analysis (PCA) aim to identify dominant trends in one dataset. However, in many settings we have datasets collected in different conditions, e.g. a treatment and a control experiment, and we are interested in visualizing and exploring patterns that are specific to one dataset. We propose a new method, contrastive principal component analysis (cPCA), which identifies low-dimensional structures that are enriched in a dataset relative to comparison data. In a wide variety of experiments, we demonstrate that cPCA with a background dataset enables us to visualize dataset-specific patterns missed by PCA and other standard methods. We further provide a geometric interpretation of cPCA and strong mathematical guarantees. An implementation of cPCA is publicly available, and can be used for exploratory data analysis in applications where PCA is currently used.

Major advances in computer science and artificial intelligence, in particular "deep learning", are beginning to have an impact on radiology. There has been an explosion of research interest and number of publications about the use of deep learning in radiology. In this presentation, I will show examples of how deep learning has led to major improvements in automated radiology image analysis, especially for image segmentation and computer aided diagnosis. I will also show how the radiology report can be used to do bulk annotation of images for training the deep learning systems.

Lung cancer remains the most common cause of cancer death worldwide and many patients will develop recurrence or progression following primary treatment. More accurate clinical, imaging and molecular biomarkers would be extremely useful for stratifying patients who are at a higher risk of recurrence and who might benefit from adjuvant or more aggressive treatment options. Maximum standardized uptake value (SUVmax) on positron emission tomography (PET) imaging has also been shown to predict recurrence or death in NSCLC. However, we hypothesized that applying a radiomics approach to extract more complex information (e.g., texture) from standard medical images could provide additional prognostic information. In this presentation I will discuss our current work on using tumor and non-tumor PET radiomics to augment traditional metrics of response in non-small cell lung cancer.

Metabolic reprogramming of the tumor microenvironment is recognized as a cancer hallmark. In this presentation, I will discuss our work on an integrative analysis of primary non-small cell lung cancer (NSCLC) with data obtained on medical imaging (¹⁸FDG-PET and CT), tissue microarray, bulk and flow-sorted transcriptomics and survival outcomes across multiple cohorts and studies. We found that tumors with higher glucose uptake were functionally enriched for molecular processes associated with invasion in adenocarcinoma (AD) and cell growth in squamous cell carcinoma (SCC). We also identified genes correlated to glucose uptake that were predominately overexpressed in a single cell-type comprising the tumor microenvironment. For SCC, most of these genes were expressed by malignant cells, whereas in AD they were predominately expressed by stromal cells, particularly cancer-associated fibroblasts (CAFs). Among these glucose uptake correlated AD genes expressed in CAFs, we identified gene GFPT2, a metabolic gene prognostically significant for lung adenocarcinoma at both gene and protein level. We also found that GFPT2 was correlated to proportions of ground-glass opacity (GGO), suggesting a role in lung adenocarcinoma histologic classification and progression. Our work provides new evidence of histology-specific tumor-stromal properties associated with glucose uptake in NSCLC using an integrative analysis, and identifies GFPT2 as a critical regulator of tumor metabolic reprogramming and progression in AD.

During this lecture, Dr. Movson will...

• Describe how a categorization system for radiology results (RADCAT) was developed, and then deployed system wide at  Lifespan and show how it was integrated into the Radiant critical results module”
• Describe an automated system for providing follow up of pathology/cytology/endoscopy/operative notes via email.

Radiology reporting is a signifiant part of the radiology workflow and play a central role in the documentation and communication of radiology findings. While existing efforts on integrating artificial intelligence (AI) systems into radiology are largely focused on image analysis, we argue that it is of equal importance to optimize the reporting system with AI algorithms. However, this is difficult due to the natural language format of radiology reports. In this talk, I will discuss two recent efforts that aim at optimizing the radiology workflow by understanding these reports. First, I will talk about a deep learning-based algorithm that learns to reliably detect urgent radiology findings from free-text reports, and therefore can trigger alerts to physicians when urgent follow-up is required. Second, I will introduce a natural language sequence-to-sequence learning algorithm that learns to summarize radiology findings into impression statements. We showed that this system is able to automatically generate impressions which are natural and clinically valid under radiologist evaluation, and has the potential to speed up existing radiology workflow.

Clinical ultrasound is relatively inexpensive, widely available, and noninvasive. In addition, due to the volume of images obtained during the course of an ultrasound examination, it is a valuable source of data for quantitative analysis, particularly using machine learning and deep learning. This talk discusses two potential applications. Ultrasound elastography is a noninvasive way of evaluating fibrosis, or tissue stiffness. This talk discusses methods to make use of elastography data in novel ways for improved disease diagnosis. In addition, molecular ultrasound has emerged as a new method of early cancer detection, which is crucial as early cancers are more treatable with a better prognosis. This talk discusses fully automated techniques for analyzing such data.