I will present new computer vision algorithms to learn complex morphologies and phenotypes that are important for human diseases. I will illustrate this approach with examples that capture physical scales from macro to micro: 1) video-based AI to assess heart function (Ouyang et al Nature 2020), 2) generating spatial transcriptomics from histology images (He et al Nature BME 2020), 3) and learning morphodynamics of immune cells. Throughout the talk I'll illustrate new design principles/tools for human-compatible and robust AI that we developed to enable these technologies (Ghorbani et al. NeurIPS 2020, Abid et al. Nature MI 2020). 

Precision medicine is an approach to patient care that considers individual differences in a patient’s genetic and molecular makeup to predict disease progression and optimize treatment response – and one of the greatest opportunities, and challenges, in modern cancer care. There is currently no known method to predict the metastatic potential of any cancer at early stages. We propose to address this critical barrier by creating a cancer digital twin, a digital replica of a cancer patient using state-of-the-art artificial intelligence (AI) techniques applied to 900,000 abdominal CT scans from a high-volume cancer center. Predicting metastatic progression at early stages would radically transform our approach to cancer treatment, but promises substantial implications for patients and society that must be considered. For example, would you want to know a dismal prognosis predicted by your digital twin? How would you act on this information – would you regard the choice as yours, or your fate as given by your twin? In a world where AI is increasingly biased, how do we ensure that AI doesn’t create further inequities? Dr. Simpson will present new work on the development of a cancer twin for predicting metastatic progression as well as provide some discussion of the social ramifications of such technology.

Machine learning advances in the past decade have revolutionized computer vision, and have since also made their way into medical imaging. Although many associate medical imaging with radiology first, clinical pathology might be more significantly impacted, at least in the short term by machine learning. Partly due to the simultaneous advent of machine learning in medicine and digital pathology, but perhaps mostly because in pathology achieving inhuman feats is surprisingly low-hanging fruit. In this presentation, I will give some examples analog to radiology, then illustrate some superhuman applications. Subsequently, I will highlight our current limitations and possible future research directions.

Advancements in machine learning and deep learning techniques carry the potential to make significant contributions to healthcare, particularly in fields of diagnosis, prognosis, and treatment decisions. However, most of current state-of-the-art machine learning and deep learning models for healthcare applications consider only a single input data stream without data informing clinical context. The trend of ignoring clinical contextual information is particularly prominent when dealing with the diagnosis and prognosis tasks where the imaging data is accessible. Yet in practice, pertinent and accurate non-imaging data based on the clinical history and laboratory data enable physicians to interpret imaging findings in the appropriate clinical context, leading to a higher diagnostic accuracy, informative clinical decision making, and improved patient outcomes. To achieve a similar goal using machine learning and increase the physician trust, clinical diagnosis and prognosis models must also achieve the capability to process contextual clinical data from electronic health records (EHR) in addition to pixel or other sensor data. This talk will present multiple fusion machine learning models on the imaging data with boosted performance by integrating the clinical context. In addition to imaging, I will also present a smart flexible sensor patch with on-chip AI capability that can be used in homecare to generate advance alert of cardiovascular abnormality by combining physiological signal data with patient demographic and comorbidity information.

Over the past decades, diagnostic imaging has become an integral part of modern medicine. Current increase in the number and volume of the acquired medical images has led to a tremendous increase in the expert workload for image interpretation. In recent years, deep learning has revolutionized many fields including medical imaging. It has shown potential to automate the routine analysis thereby alleviating expert workload, as well as to support clinical research. In this presentation, I will show our recent work on the development of AI-based methods to improve image quality, I will present methods we developed for quantitative analysis of CT and MRI exams, and illustrate their application in large-scale studies.

 Artificial intelligence, in particular from the class of machine learning and deep learning, has shown great promise for application in medical imaging. However, the success of AI-based techniques is limited by the availability and quality of the training data. A common approach is to train methods on well annotated and curated databases of high-quality image acquisitions, which then may fail on real patient cases in a hospital setting. Another problematic is the lack of sufficient numbers of clinical label annotations in the training data, or example for early markers of disease. In this talk I will present some of our recent approaches that aim to address some of these challenges, by using AI as an enabling technique for improved image reconstruction, realistic data augmentation and further downstream tasks. I will conclude by giving an outlook on the future opportunities in this field, operating right from the imaging sensor to extracting clinically relevant measures.

In this talk, I will present methods for future cancer risk from medical images. The discussion will explore alternative ways to formulate the risk assessment task and focus on algorithmic issues in developing such models. I will also discuss our experience in translating these algorithms into clinical practice in hospitals around the world.

Diabetic retinopathy is the leading cause of blindness in working age adults in the United States. It is challenging to address in both rural and urban underserved settings, which suffer from shortages of eye specialists. This talk will describe the approach taken to address this condition in a medically underserved area (South Los Angeles) by researchers in the Center for Biomedical Informatics at Charles R. Drew University of Medicine and Science, using telehealth and machine learning on data from patient electronic health records.

With the recent expansions of whole-slide digital scanning, archiving, and high-throughput tissue banks, the field of digital pathology is primed to benefit significantly from deep learning technology. This talk will cover several applications of deep learning for characterizing histopathological patterns on high-resolution microscopy images for cancerous and precancerous lesions. Furthermore, the current challenges for building deep learning models for pathology image analysis will be discussed and new methodological advances to address these bottlenecks will be presented.

Automated methods for detecting prostate cancer and distinguishing indolent from aggressive disease on Magnetic Resonance Imaging (MRI) could assist in early diagnosis and treatment planning. Existing automated methods of prostate cancer detection mostly rely on ground truth labels with limited accuracy, ignore disease pathology characteristics observed on resected tissue, and cannot selectively identify aggressive (Gleason Pattern≥4) and indolent (Gleason Pattern=3) cancers when they co-exist in mixed lesions. This talk will cover multimodal and multi-scale fusion approaches to integrate radiology images, pathology images, and clinical domain knowledge about prostate cancer distribution to selectively identify and localize aggressive and indolent cancers on prostate MRI. 

Transfer learning has quickly become standard practice for deep learning on medical images. Typically, practitioners repurpose existing neural networks and their corresponding weights to bootstrap model development. This talk will cover several methods to pretrain neural networks for medical tasks. The current challenges for pretraining neural networks in Radiology will be discussed and recent advancements that address these bottlenecks will be highlighted.