Open-Source Three-Dimensional Digital Pathology: Application to Serially Sectioned Renal Needle Core Biopsies
Ashley Ward, Seymour Rosen, Charles Law, Beverly Faulkner-Jones. Beth Israel Deaconess Medical Center, Boston, MA; Kitware Inc., Clifton Park, NY
Background: Evaluation of microscopic three-dimensional (3D) structure of tissue has great potential in anatomic pathology. Currently evaluation of serial sections to generate a “3D” data set, e.g. medical renal biopsies, are performed manually on 2D tissue sections from which the 3D structure is inferred. With advances in whole-slide imaging (WSI), pathologists now have the opportunity for digital 3D analysis. The renal parenchyma is complex and geometric, and stratified and compartmentalized at all levels. A digital 3D construct would provide insights difficult to appreciate on 2D views.
Design: We have developed software tools for aligning WSI of serial tissue sections, and for viewing the resulting large-scale 3D image datasets. These tools are integrated into a web-based digital pathology system that will be released open source. The alignment algorithm differs from existing solutions in that it is designed to handle the artifacts typical of tissue sections including warping, folding, tearing and loss of tissue. To minimize needed server resources, field transformations are stored with the original (unmodified) image data and applied using a WebGL viewer. The system uses a load-on-demand strategy and progressive rendering to ensure the WebGL viewer remains responsive over standard networks. With this approach users are able to interactively navigate large volumes (i.e., quickly step through sections, pan, zoom and rotate).
Results: We have generated 3D image volumes from serial sections of renal core biopsies. Here we present an example from one full biopsy sectioned at ∼2 microns to generate 66 tissue profiles on 23 slides (2 to 4 tissue profiles / slide). Sections were first roughly aligned and then finely aligned (Figure 1). The aligned whole slide images were then annotated and individual components measured and evaluated.
Figure 1: For fine alignment glomeruli are detected using texture classification
Conclusions: Using our software and alignment algorithm we are able to generate 3D image volumes from routinely processed renal biopsies. This shows the feasibility of 3D image modeling from 2D images from any part of the renal parenchyma. Examples will be shown.
Monday, March 4, 2013 11:30 AM
Proffered Papers: Section H2, Monday Morning