Digital Pathology and its importance as an omics data layer

Yves Sucaet [1] [2]

Bioinformatics and pathology are obvious scientific partners. Bioinformatics often takes places at the most basic (almost chemical, or even physical) level of life, but much of its procedures to obtain data are destructive. Pathology on the other hand takes place at a much more coarse level of data acquisition (usually where the physical properties of visible light end), but has the advantage of being rooted in the tradition of medicine. The traditional paradigm of pathology is "tissue is the issue". Morphology (exactly the component that often gets overlooked in bioinformatics) plays a large role and helps millions of patients each year around the world. Pathology is proven technology, bioinformatics is limited to niche applications.
With the development of whole slide imaging technology some twenty years ago, digital pathology became possible. Observations that used to be for the eyes of the pathologist only, could now be captured and translated into high-resolution pixels, and studied by and communicated to many. Many began to dream of automated tissue evaluation systems and AI-pathology, some even going as far as to suggest the replacement of the pathologist by intelligent computer systems.
Meanwhile in several areas of bioinformatics, new limits are being hit. Yes, we can do high-throughput experiments, but noisy datasets are often the results, (inter- and even intra-observer) replicability is difficult, and statistics only offer limited relief.
The goal of this introductory lecture is to highlight the problems as well as opportunities for both fields of study, and how exchange of experiences, and (in a later stadium) integration of techniques close the scientific gap that still exists in a great many areas.
There is no lack of pathology-centric workshops that offer insights into the world of algorithms. With the CPW event however, we take another approach. We want to bring together the most advanced groups in digital pathology, with the bioinformatics community, to explore the opportunities that exist on both sides of the fence.
We start by explaining the basic data types that are introduced by digital pathology. We also explain where they come from, and why this presents unique challenges when it comes to data mining and image analysis. Finally, we introduce PMA.start, a free software environment that can be used to universally gain access to digital pathology (imaging) data.
Bioinformatics groups can help quantify, model, and reduce morphological whole tissue data. Pathologists can help interpret and explain heterogeneous high-throughput datasets. And the first seeds of such collaboration can be planted right here, in Athens.

Keynote lecture: Fifteen Years of Digital Pathology at Leeds: The Journey To Clinical Adoption of Digital Pathology and Where Now?

Derek R. Magee [1]

It is fifteen years since the first digital pathology scanners were installed in the pre-clinical lab in Leeds. This talk charts our journey from the installation of those first scanners, through our pre-clinical image analysis, and interface design, research, to current clinical adoption of the technology in the hospital. Along the way we have built our fair share of software, collected hundreds of terabytes of images, created the largest digital pathology teaching resource in the world (virtualpathology.leeds.ac.uk), generated hundreds of 3D pathology datasets, and set up a spin out company (HeteroGenius Ltd.). The talk will also touch on where we go from here. Now we have fully digital workflows how can image analysis and AI be brought into the clinic? How can massive volumes of clinical data facilitate pre-clinical research without compromising patient confidentiality?

Towards the complete picture: computational strategies to identify MS-Imaging derived molecular fingerprints associated with inflammatory cell phenotypes

Chang Lu1 [1], Pieter Goossens [1], Joel Karel [2], Marion Gijbels [1], Evgeni Smirnov [2], Benjamin Balluff [3], Ron Heeren, [3], Erik A.L. Biessen [1]

Background: Atherosclerosis is the leading cause of coronary heart disease, cerebral infarction, and peripheral vascular disease. Macrophages are involved in the pathogenesis of atherosclerosis and play a significant role in its occurrence, progression, and regression. They were recently shown to be very plastic cells, able to adapt to stimuli in their direct microenvironment, to adopt a spectrum of functional phenotypes (e.g. M1, M2, Mφ) throughout the disease course. The heterogeneity in plaque is only poorly described, largely due to technical constraints. Here we present a powerful new strategy to dissect macrophage heterogeneity in its molecular context at high spatial resolution.
Methods and results: Hereto, we analyzed plaque sections by multispectral imaging microscope to concomitantly visualize up to 12 biomarkers in one single section. Images were deconvoluted to the separate biomarker components (features). We next defined cell segments in the image using a watershed cell detection method (Qupath software), extracted the feature intensities for each cell segment within the plaque and constructed a feature intensity matrix. Based on the feature intensities, we clustered the cell segments using K-means, DBscan, and Self-organizing feature Map (SOM), with different distance measures (Euclidean, cosine, etc). We identified 20 different cell clusters for all of the clustering methods. Image inspection by a mouse pathologist indicated that cosine similarity based SOM (coSOM) was performing best. Next, we set out to map the molecular context on an adjacent section, using Mass Spectrometry Imaging (MSI). The two sections were aligned automatically, based on morphological landmarks, by an elastic registration algorithm mapping the positions of the identified macrophage clusters on the MSI dataset. In the near future, we expect to select crucial MSI features that are most characteristic of a particular cell cluster, from the tens of thousands of m/z peaks by sparse Partial Least Squares Discriminant Analysis (sPLSDA), a robust feature extraction method. M/z peaks from the MSI feature list corresponding with a particular macrophage cluster will be assigned to a specific chemical compound.
Conclusion: We have developed an innovative new approach to define not only cell heterogeneity, but also to link this to the microenvironment. The methodology is also applicable to other cell types and tissue and may lead to a breakthrough in linking molecular context to cellular phenotype and function, in healthy and diseased tissues.

Keynote lecture: Computational Systems Pathology and Machine Learning for Mechanistic Understanding of Cancer Metastasis

Chakra Chennubhotla [1]

Determining the mechanistic underpinnings of progression in metastatic disease is a major unmet need. Metastases involves the local migration and dissemination of cancer cells from the primary tumor and their survival, albeit with high attrition rates, leading to lethal secondary tumors at distant sites in vital organs. Comprehensive genetic profiling has revealed intrinsic molecular variability, or intratumor heterogeneity (ITH), in multiple cancers. Additionally, multiple lines of evidence across several tumor types indicate that spatial, functional, and genomic intratumor heterogeneity (ITH) among malignant cells, non- malignant cells (e.g., immune cells, cancer associated fibroblasts (CAFs), endothelial cells), and their localized interactions within the tumor microenvironment (TME) is a critical determinant of disease progression landmarks that include metastasis, immune evasion, therapeutic response, and drug resistance.
With the goal of understanding the transition from locally invasive to metastatic cancer, I will describe our progress in building THRIVE (https://ith.csb.pitt.edu/), a comprehensive iterative computational systems pathology platform employing in situ immunofluorescence (IF) labeling and imaging technology and novel machine learning algorithms. THRIVE is capable of deciphering diverse molecular and cellular signaling networks conferring malignant (metastatic) phenotypes within distinct sub-regions. Systems analysis of these interactions provides a network model predicting additional (emergent) interactions which can then be tested iteratively on the same tissue sample. Corroboration of the predictive model will generate causal hypotheses that could inform novel therapeutic strategies to remodel the primary tumor TME to be less permissive for metastasis.

Keynote lecture: Computational Pathology at Scale: Changing Clinical Practice One Petabyte at a Time

Thomas Fuchs [1], [2]

Pathology is in the midst of a revolution from a qualitative to a quantitative discipline. This transformation is fundamentally driven by machine learning in general and computer vision and deep learning in particular.
At Memorial Sloan Kettering we are building a computational pathology decision support system based on hundreds of GPUs and one petabyte of image and clinical data. The goal is not only to automated cumbersome and repetitive tasks, but to impact diagnosis and treatment decisions in the clinic.
This talk will focus on our recent advances in deep learning for tumor detection and segmentation, on how we train these high capacity models with annotations collected from pathologists and how the resulting systems are implemented in the clinic.