Patient Similitude: Combining multiple-scale imaging & molecular tissue and other phenotypes

NII Shonan Meeting:

@ Shonan Village Center, November 12 – 15, 2018


          • Raghu Machiraju, The Ohio State University, USA
          • Jens Rittscher, University of Oxford, UK
          • Motonori Ota, Nagoya University, Japan


Description of the meeting

Patient data is increasingly available in many forms including genomic, transcriptomic, epigenetic, proteomic, histologic, radiologic, and clinical. Repositories such as The Cancer Genome Atlas (TCGA) contain multitude of patient records which can be used for patient stratification, disease subtype/biomarker discoveries, all essential elements in the era of personalized medicine. Still, patient records from the clinic are used singularly to diagnose patients in the clinic without including all likely insights from other sources. Similarly, molecular expression signatures from the omic sources do not significantly impinge on the clinical best practices.

There is a need to bridge the gap between the laboratory and the clinic. The need is even more urgent given the need to assess outcome to treatment. Latest therapeutic advances promise to extend lives of patients by years rather than just weeks or months. An example for such a highly promising strategy is cancer immunotherapy. But today we do not understand why certain patients benefit from a given therapy while others do not respond.

Fundamental methodological advances are necessary to achieve this goal. While genomic information provides important a priori information on potential risk profiles, gene expression and transcriptomic data tells us which genes are activated to likely produce salient proteins, metabolomes and other phenotypic traits. Epigenetic, transcriptional regulation, and post-translational modifications can manifest in specific patterns at the tissue scale and often captured as histology images. Studies on large patient cohorts, such as genome wide association studies, can only provide partial answers. The stratification of patients into subgroups that not only have a similar genomic risk profiles, but also show the same therapeutic response is a question we cannot answer with existing approaches. Still quantitative data derived from images taken at the cellular and tissue level has not played the role that it potentially can. It can play a crucial role in aiding the necessary diagnoses.

We submit that the association of image based phenotypes with genotypes is necessary for identifying disease subtypes that are consistent with our molecular understanding of disease. Imaging provides important complementary context information on multiple different anatomical levels. Thanks to advances in medical image analysis, computer vision and machine learning we have made tremendous progress in analyzing medical images acquired on CT, MR, PET or ultrasound machines. However, histology images have not benefitted from a systemic studies given the larger size and complexity of tissue level imaging. Enabled by the more widespread availability of digital slide scanners the field of computational pathology can be furthered by new scrutiny.

This workshop provides a unique opportunity to bring together clinical scientists, word leading experts in bioinformatics as well as computer scientists and biomedical engineers to discuss these challenges and nucleate a scientific community that will address these problems in a highly structured and systematic way. We especially focus on creating clean measures of patient similitude and in turn develop workflows and machine learning methods that result in interpretable outcomes in the clinic and most importantly use multiple kinds of data. Where necessary we will invoke the need for deep learning techniques (image processing of histology images) and integrative methods based on Grassmannian manifolds. We contend that the full potential of these technologies on public health critically depends on our ability to systematically link and integrate patient information in the form evidence and risk. We contend that the full potential of these technologies on public health critically depends on our ability to systematically link and integrate patient information in the form evidence and risk.

Principal informatics challenge
Evidence-based methods that point to a multidimensional and multifactorial definition of patient similarity or patient similitude is certainly the need of the hour. The following three questions will be used to establish a structure for the workshop:

How can we make the transition from categorical disease descriptions to representations that will foster the development of evidence-based notion of patient similarity?

Through an integration of multi-scale information that captures imaging as well as molecular and genetic data we hope to build representations that can effectively stratify the patient population in terms of disease outcome. Integrated clustering methods have already been explored but need to be extended to capture imaging information.

How does one measure patient similarity?

No existing standard similarity metric will suffice the requirements for measuring patient similarity. Here it will not only be necessary to capture image based features from different anatomical scales but also take genetic and other molecular stratifiers into account. Today, clinical experts intuitively incorporate clinical best practices, personal experience as well as a limited amount of sparse and
interpretable diagnostic tests. We need to explore new paradigms for constructing metrics that are capable of incorporating such information but are founded in evidence provided by the available patient data.

How can existing practices of image analysis, computer vision, machine learning and visualization be leveraged?
With the recent advent of deep learning, there has been interest in exploring intrinsic features in large histology images. These intrinsic features can represent both large-scale and fine and small-scale architectural landmarks and artifacts. However, they are difficult to obtain. Extrinsic features which have been the staple of computer vision and image analysis although easy to realize are more difficult to identify in a histology image. Can one find hybrid solutions? Further, are there viable correlative methods that can be used to associate different phenotypic signatures with either intrinsic or extrinsic features.

To capture the various aspects of each of the questions outlined above we will make apply the following elements to discuss and explore the various aspects of each topic:

  • Keynote presentations (to be determined)
  • Spotlight presentations by young scientists,
  • Question led discussion groups (led by two senior members).

Expected outcomes of the meeting:

  • Connect researchers who aim to address these challenges but are currently working in different scientific communities
  • Provide a forum or young scientists who are interested to work in this space
  • Develop the foundation for a high-profile series of symposia and workshops
  • Publications in form of edited volumes (to be discussed)

Comments are closed.