Magnetic resonance imaging (MRI) is an increasingly important tool for the diagnosis and treatment of prostate cancer. However, MRI interpretation suffers from high interobserver variability and often misses clinically significant cancers. Registration of histopathology images from patients who have undergone surgical resection of the prostate onto pre-operative MRI images allows direct mapping of cancer location onto MR images. This is essential for the discovery and validation of novel prostate cancer signatures on MRI. Traditional registration approaches can be computationally expensive and require a careful choice of registration hyperparameters. We present a deep learning-based pipeline to accelerate and simplify MRI-histopathology image registration in prostate cancer. Our pipeline consists of preprocessing, transform estimation by deep neural networks, and postprocessing. We refined the registration neural networks, originally trained with 19,642 natural images, by adding 17,821 medical images of the prostate to the training set. The pipeline was evaluated using 99 prostate cancer patients. The addition of the images to the training set significantly (p < 0.001) improved the Dice coefficient and reduced the Hausdorff distance. Our pipeline also achieved comparable accuracy to an existing state-of-the-art algorithm while reducing the computation time from 4.4 minutes to less than 2 seconds.

Generative adversarial networks (GANs) are advanced types of neural networks where two networks are trained simultaneously to perform two tasks of generation and discrimination. GANs have gained a lot of attention to tackle well known and challenging problems in computer vision applications including medical image analysis tasks such as medical image de-noising, detection and classification, segmentation and reconstruction.In this talk, we will introduce some of the recent advancements of GANs in medical imaging applications and will discuss the recent developments of GAN models to resolve real world imaging challenges.

During the last decade, results from basic research in the fields of genetics and immunology have begun to impact treatment in a variety of diseases. Checkpoint therapy, for instance has fundamentally changed the treatment and survival of some patients with melanoma. The medical workplace has transformed from an artisanal organization into an industrial enterprise environment. Workflows in the clinic are increasingly standardized. Their timing and execution are monitored through omnipresent software systems. This has resulted in an acceleration of the pace of care delivery. Imaging and image post-processing have rapidly evolved as well, enabled by ever-increasing computational power, novel sensor systems and novel mathematical approaches. Organizing the data and making it findable and accessible is an ongoing challenge and is investigated through a variety of research efforts. These topics will be reviewed and discussed during the lecture.

Although many radiology AI efforts are focused on pixel-based tasks, there is great potential for AI to impact radiology care delivery and workflow when applied to reports, EMR data, and workflow data. Radiology-pathology correlation, identification of follow-up recommendations, and report segmentation can be used to increase meaningful feedback to radiologists as well as to automate tasks that are currently manual and time-consuming. When deploying AI within the clinical workflow, there are many challenges that may slow down or otherwise affect the integration. Careful consideration of the way in which radiologists may expect to interact with AI results should be undertaken to meaningfully deploy radiology AI in a safe and effective way.

In this seminar, we will discuss clinical image data for machine learning.  We will explore different types of image data and data formats, tools for data annotation, and image pre-processing steps. We will also discuss inherent challenges and unique institutional experiences of data share for multi-center studies for machine learning and what we have learned from various pediatric and neuro vision projects. 

In this talk, Dr. Tiwari will focus on her lab’s recent efforts in developing radiomic (extracting computerized sub-visual features from radiologic imaging), radiogenomic (identifying radiologic features associated with molecular phenotypes), and radiopathomic (radiologic features associated with pathologic phenotypes) techniques to capture insights into the underlying tumor biology as observed on non-invasive routine imaging. She will focus on clinical applications of this work for predicting disease outcome, recurrence, progression and response to therapy specifically in the context of brain tumors. She will also discuss current efforts in developing new radiomic features for post-treatment evaluation and predicting response to chemo-radiation treatment. Dr. Tiwari will conclude with a discussion on her lab’s findings in AI + experts, in the context of a clinically challenging problem of post-treatment response assessment on routine MRI scans.

Access to large, high quality datasets collected from heterogeneous patient populations and imaging modalities is essential for researchers to explore and validate hypotheses that will generalize beyond their own institution. This is especially important as researchers increasingly apply deep learning or radiomics techniques where minor variations in the image data can lead to spurious correlations with patient diagnoses and outcomes.

Unfortunately, regulatory constraints make broad sharing of medical data from human subjects a complex process which most individual investigators and institutions are not equipped to handle. At the same time, there is increasing pressure on researchers from funding agencies and publishers to share data.  The National Cancer Institute funded The Cancer Imaging Archive (TCIA) in 2011 to help bridge this gap.  TCIA provides hands-on assistance to help investigators safely de-identify their data, as well as long term hosting to assure stable access to research community.  Each dataset is published with a title, author list, abstract, and persistent Digital Object Identifier so that they can be properly cited by other researchers and linked to their ORCiD profile.

TCIA now contains over 115 unique data collections of more than 50 million images. Recognizing that images alone are not enough to conduct meaningful research, most collections are linked to rich supporting data including patient outcomes, treatment information, genomic / proteomic analyses, and expert image analyses (segmentations, annotations, and image features). This lecture will teach attendees how to propose new datasets for publication and how to access existing datasets on TCIA.  It will also touch on common use cases for public data sharing and highlight a variety of research activities conducted using TCIA data.

The work setting of the future presents an opportunity for human-technology partnerships, where a harmonious connection between human-technology produces unprecedented productivity gains. A conundrum at this human-technology frontier remains - will humans be augmented by technology or will technology be augmented by humans? We present our work on overcoming the conundrum of human and machine as separate entities and instead, treats them as an assemblage. As groundwork for the harmonious human-technology connection, this assemblage needs to learn to fit synergistically. This learning is called assemblage learning and it will be important for Artificial Intelligence (AI) applications in health care, where diagnostic and treatment decisions augmented by AI will have a direct and significant impact on patient care and outcomes. We describe how learning can be shared between assemblages, such that collective swarms of connected assemblages can be created. Our work is to demonstrate a symbiotic learning assemblage, such that envisioned productivity gains from AI can be achieved without loss of human jobs. 

Specifically, we are evaluating the following research questions: Q1: How to develop assemblages, such that human-technology partnerships produce a "good fit" for visually based cognition-oriented tasks in radiology? Q2: What level of training should pre-exist in the individual human (radiologist) and independent machine learning model for human-technology partnerships to thrive? Q3: Which aspects and to what extent does an assemblage learning approach lead to reduced errors, improved accuracy, faster turn-around times, reduced fatigue, improved self-efficacy, and resilience? 

Dr. Callcut will review the current state of AI in Healthcare. She will also explore the hope for the future, the ethical and regulatory challenges, and the unknowns in bringing AI to Healthcare. She will provide examples of the state of the art and also discuss future compute options.

AI-based tomography is an important application and a new frontier of machine learning. AI, especially deep learning, has been widely used in computer vision and image analysis, which deal with existing images, improve them, and produce features. Since 2016, deep learning techniques are actively researched for tomography in the context of medicine. Tomographic reconstruction produces images of multi-dimensional structures from externally measured “encoded” data in the form of various transforms (integrals, harmonics, and so on). In this presentation, we provide a general background, highlight representative results, and discuss key issues that need to be addressed in this emerging field.