Quantitative Tissue Algorithms

Thursday, October 6, 2011
Jason Hipp, MD, PhD
“Quantitative Histopathology Analysis”
The increasingly prevalent ability to create digital whole slide imaging (WSI) data sets (digital slides) from glass slides offers new opportunities to quantify aspects of morphology and histopathology. In addition, the adoptions of all digital pathology workflows in the near future will create millions of digital slides, and here I will discuss digital tools that may help extract quantitative information from this data mine. 
Spatially invariant vector quantization (SIVQ) is a recently described pattern recognition algorithm that can identify morphologic and architectural features on H&E and special stained digital slides [1]SIVQ differs fundamentally from other image analysis algorithms in its easy to use and zero training workflow.  Initially, one (or a small number of) distinct predicate image candidate feature(s) is identified by the user. This feature is then utilized to perform an exhaustive search of the entire surface area of the digital slide, resulting in the generation of a statistical probability heatmap of quality of matches. Introductory video demonstrations of typical uses of the SIVQ algorithm are available online [2]
            I will discuss new features that have been added to the SIVQ algorithm, such as a fully automated selection process that yields optimal vectors without iterative discovery. Another new feature is complete rendering of the SIVQ algorithm within a parallel computation environment, via the use of the MATLABTM API. Together, these two innovations render the SIVQ algorithm more suitable for actual clinical deployment by virtue of newly-gained automation and speed, respectively. Lastly, I will discuss an additional feature that enables the calculation of highly accurate, precise surface area measurements of morphologic features to two significant digits and its application to clinical use cases.  
I will also discuss user-friendly tools for pathologists that enable pathologists to mine digital slides archives to create image microarrays (IMA). IMAs are to digital slides like Tissue Microarray (TMA) are for cell blocks. Thus, a single digital slide can contain up to hundreds-thousands of high quality images containing key diagnostic morphologies and appropriate controls (benign mimics and background features). Representing hundreds to thousands of adjudicated features (representing diverse sets of morphologic variants) improves the efficiency of screening image analysis/pattern recognition algorithms much quicker and computationally less computationally.
Current methods for cutting and cropping out features from digital slides is done by dragging open a window which makes creating images of the same size and resolution extremely challenging and tedious. We therefore designed a tool that enables the systematic capturing of image features at the same size and resolution. This tool has numerous potential applications to clinical and research workflow. 
Lastly, I will discuss a WSI repository for publications (www.wsirepository.org) that we have created to provide access to the entire digital slides rather than limited fields of views for published images, similar to what the microarray community did with raw gene expression data [3]. We believe this will increase the level of rigor for peer review, allow for multiple perspectives to generate potentially alternative interpretations, improve educational opportunities, and provide reference data sets for algorithm development and inter-observer concordance.

Copyright 2009 | Site by UMC Web Team