First, we will briefly review the history of informatics in radiology. Three past eras are identified, and the requirements for the new era of quality, value, and personalized or precision medicine are described. The challenges of proving the value contribution of imaging to patient care through informatics are identified. Finally, the new expectations of collaboration with our clinical colleagues within the new models of healthcare reimbursement are described, from the context of an academic radiology department.

Dr. Erickson will describe their recent work applying traditional machine learning methods to brain tumors, for the purpose of assessing therapy effects, selecting best therapy options, and identifying genomic/epigenomic properties of tumors. Those results will then be compared with early investigations into deep learning for the same tumors and outcomes. Finally, he will describe the computing system they have developed to make such investigations more efficient.

Much research work has been done on medical information analysis and medical image retrieval. In the past few years the used data sets have increased in size and complexity and medical big data has become an important topic. Scientific challenges have also increased and gained much impact in several fields, as they allow comparing algorithms based on the same data and scenario and thus to really measure progress. Still, problems remain when data are too large to be distributed, contain confidential information and when real time data are needed. The Evaluation as a Service (EaaS) paradigm aims at creating research infrastructures that solve some of these challenges and working towards a sustainable biomedical big data infrastructure.

Cancer development is a highly complex evolutionary process driven by the dynamic feedback between somatic mutations and the organ microenvironment in which the tumor initiates and progresses. Spatial and temporal gradients of microenvironmental factors (e.g. nutrients, oxygen, and drug) result in selective pressures that affect the fitness of cancer cells and influence the overall composition of the tumor. It is imperative to capture the heterogeneity of a tumor across scales (of size and time) to achieve a deeper understanding of the evolutionary dynamics of a tumor and the ability to successfully intervene therapeutically. Recently, we developed a high-content imaged based method that is capable of multiplex phenotypic profiling of cell populations in response to co-occurring environmental selective pressures. This platform has been applied to both 2D monolayers and 3D patient-derived organoids to track changes in cell/organoid growth and morphology features under various environmental perturbations. Here we demonstrate that the local tumor microenvironment can influence the magnitude of the selective effect, with larger and more numerous organoids forming under physiologically relevant oxygen culture conditions (5% and 1% O2) compared to standard laboratory conditions (21% O2). Our image-based approach provides a deeper insight into the evolving cellular dynamics of a tumor, and utilizes reduced reagents and time, offering advantages over traditional biological assays.  

A cell’s stimulus response depends not only on the internal concentrations of signaling molecules but also on their spatial dynamics. Thus we need techniques to measure the bulk kinetics of molecules in live cells to better identify mechanisms responsible for cell behavior. In this talk I will present a new method called multimodal Image Correlation Spectroscopy (mICS) that estimates ballistic and anisotropic diffusive molecular motion from live cell microscopy. I will demonstrate the utility of mICS as applied to a prostate cancer (PCa) cell line system. Androgen receptor (AR) is the key PCa signaling molecule that drives proliferation by translocating from cytoplasm to nucleus under ligand stimulation. This signal is the main therapeutic target in early stage PCa. mICS analysis shows that AR undergoes diffusion, confined movement, and bound motion following ligand stimulation. Together with and simulations from a simple computational model, these new findings imply that cytoplasmic AR translocation is dominated by diffusion rather than microtubule transport. This has implications for therapeutic strategy.

Small pulmonary nodules are very common and found on chest CT in about 50% of adults over 50 years of age. Although the vast majority of these nodules are benign, a small proportion is due to potentially curable lung cancer, which is the leading cause of cancer death in the US. The complex multistep process of CT detection, risk assessment, radiologic reporting and patient follow-up is time consuming, labor intensive and highly variable among radiologists and medical institutions. However, the process is ripe for streamlining and standardization with the recent development of the ACR Lung Imaging Reporting and Data System (Lung-RADS™) for screening and the imminent revision of Fleischner Society guidelines for incidentally detected small pulmonary nodules. In this presentation, I will provide some relevant background on small pulmonary nodules, describe the complex multistep process in current practice and highlight opportunities for improvement with applications of imaging informatics.

Over the last thirty years, neuroscience instruments have grown massively in their sophistication; the number of software tools to analyze the data has grown commensurately.  The increase in data size and computational complexity poses a challenge for individual scientist and for projects that integrate the work of multiple scientists. The complexity has reduced the transparency of published results, reducing confidence in scientific publications.  Victoria Stodden suggests that the core of the problem is the difficulty complexity creates when try to check one another’s work.

When Bob Dougherty and I developed the infrastructure for Stanford’s Center for Cognitive and Neurobiological Imaging (CNI), we were determined to try to help investigators manage the large scale neuroimaging data and to support them by providing computational tools.  Our ambition was to provide an experimental environment that supports the ability to check one another’s work. 

Based on this aspiration, a small group of us started the Project on Scientific Transparency. I will describe how we see the problem in neuroimaging, though I suspect most of you know the issues all too well. Then I will describe the tools we are developing to help neuroimaging scientists to check each other’s work. The informal presentation will include time to discuss the pragmatics of how decisions are being made and to listen to your feedback about future directions.

Actionable focal masses representing possible cancer are commonly detected on imaging exams; these findings are present on abdominal CT scans in up to 37% of the well population and 31% of patients with cancer. Yet radiologists variably describe the malignant potential of these masses in reports and inconsistently provide follow-up recommendations. This contributes to delayed or missed cancer diagnoses due to a) provider misinterpretation of the likelihood that a mass represents cancer and b) poor or absent monitoring systems within free text reports. Additionally, correlation of imaging findings with other outcomes, including health care utilization, costs and cancer diagnoses are challenging within traditional free text reports.

Our health system utilizes a standardized assessment coding scheme for reporting the malignant likelihood of masses in the abdomen and pelvis similar to the system used in mammography; codes are included in all abdominal CT, MRI and ultrasound reports regardless of the indication for the study or patient location (i.e. outpatient, inpatient and emergency department). This talk will discuss why and how this system was created and how codes can be used to measure differences between radiologists and to eventually correlate imaging findings with patient outcomes.

Machine learning has demonstrated great potential in radiology, oncology, ophthalmology, pathology and many other domains in medicine in diagnosis, grading and assessment of response to therapy. Supervised machine learning requires large amounts of well-annotated training data, which can be hard to obtain. Further, the lack of agreement among experts can lead to biased and noisy training datasets. I will present our recent work exploring potential reasons for lack of agreement amongst experts in rating images for Retinopathy of Prematurity (ROP) and the impact this has on supervised learning.  I will also describe the evolution of techniques in our journey to create an automatic screening tool for ROP from more traditional image processing to machine learning to deep learning. Deep learning, facilitated by advances in hardware, software and algorithms, has emerged as a leading technology in computer vision and image analysis and is being applied to medical imaging. However, deep learning requires large amounts of data, which can be hard to obtain. Archives such as the Cancer Imaging Archive (TCIA) are great resources especially for researchers without access to internal clinical data, and have helped move the field forward by making available large amounts of publicly available data. Using the Brain Tumor Segmentation (BraTS) Challenge as an example, I will briefly discuss the advances in performance of segmentation algorithms for gliomas over the last 5 years and the potential roles of automatic segmentation in large scale radiogenomics analysis.