In recent years, we have focused on developing robust general image analysis methodology, culminating in our pattern recognition tool called WND-CHRM. We have validated this pattern-recognition approach to biological image analysis using diverse imaging modalities ranging from fluorescence microscopy to X-rays of human knees, MRIs, and other imaging modalities. We have also validated a range of applications from scoring image-based assays to diagnosis of disease to prediction of future disease risk. The specific applications of this approach are covered in reports AG000674-12 and AG000685-09. A major effort recently has been to rewrite the WND-CHARM code-base to make it more modular, better organized, easier to use, and accessible with the Python scripting language. A recent release of WND-CHARM 1.52 is available from our code repository (http://ome.grc.nia.nih.gov/wnd-charm/wndchrm-1.52.775.tar.gz). This release focuses on streamlining access to image feature computing algorithms so that they are individually acessible by the name of the algorithm and any preceding image transforms. The individual algorithms are also broken out behind a common interface so that new algorithms are easy to add. The image feature computing plan was previously hard-coded, but to make it more flexible, easier to modify, and responsive to on demand feature requests, it is now automatically assembled based on the requested features, taking account of their preceding image transform dependencies. Spatially-resolved pattern analysis places an extreme burden on the performance of our software. Instead of an entire image being considered at once, or split into a small number of tiles on a grid, to achieve spatial resolution, each image must be sampled thousands or millions of times. In order to make this type of application practical, the computational strategy used in the software must be reconsidered. Previously, all 3,000 low-level image features were calculated for each image sample, even when most of them were later found to be irrelevant to the classification problem because they lacked discrimination power. The major change in strategy to enable spatially-resolved pattern recognition is to eliminate unnecessary calculations. This requires an on-demand computing strategy for image features, which is a major architectural change in the wndchrm software. Our current release makes use of this strategy, exposing a very simple to use API for specifying the specific features to be computed. An added benefit of this restructuring is that it is now very straightforward to add new image descriptors to this library. Because all of the existing descriptors use the same interface, they act as examples for future additions. We have have also made the majority of the underlying C++ code accessible from the Python scripting language to make it easier to customize how WND-CHARM is used in new applications. It is now possible to compute on-demand features using the Python interface and perform further processing using mathematical, scientific, and machine learning libraries available for Python (numpy, scipy, Scikit-learn). A major feature expanded in the past year is development and refinement of the sliding window classifier. This feature allows an entire image to be scanned systematically to look for the presence of patterns that the classifier has been previously trained to recognize, either by defining them manually, or with the use of automated segmentation methods. A second, related feature is the ability to train a classifier to isolate objects of interest using a segmentation mask. One or more binary images defining segmentation masks are automatically randomly sampled with a user-defined window size to define image-windows that are within the desired object, constitute a boundary, or are outside of the object of interest. Subsequently, a new image is systematically scanned using a sliding-window classifier to define the boundary and/or interior of an object of interest. In 2012, in collaboration with Jason Swedlow (University of Dundee, Scotland), a large international project sponsored by the Wellcome Trust was initiated to develop specific applications of the OME/OMERO system. Our group's contribution to this project involves providing interfaces between OMERO and WND-CHARM to enable image comparisons in large, diverse image repositories. The eventual goal is to use pattern recognition to annotate new images added to these collections automatically, based on previously annotated images and a large set of independent classifiers that opeerate autonomously in the background. The primary design goal of a system like OME/OMERO is to provide scientists with easy ways of organizing and annotating their image collections. The organizational structure of the images and their grouping by their annotations can also serve as the primary inputs for training pattern-recognition classifiers. Because classifiers require little or no additional input from the user, the natural convergence of these two technologies represent a powerful new mode for maximizing the utility of large scale scientific and medical image databases. Currently we have a functioning prototype that interacts with OMERO to read image data and annotations; use these for training a classifier; and return annotations derived from classification back to OMERO. Substantial work remains to make this integrated system practical. Currently, a developer-preview release is available, but it lacks the flexibility on the OMERO side to keep track of multiple, potentially conflicting classifications for images, as well as maintaining a flexible store of image feature sets. Our current efforts are to implement a standardized feature store based on the widely used HDF data format, which can be used both standalone with WND-CHARM alone, and when using WND-CHARM together with the OME database. We are also populating this database with a large and diverse set of images from cell-based high-content screening assays, as well as their corresponding CHARM features. This will constitute a realistic test-case for computation, storage and retreival of images and their corresponding numerical image descriptors computed with our software. Several efforts are under way to use parts of the OME Data Model in Pathology Informatics. Pathology makes use of digital microscopy and the Pathology Informatics field is concerned with the standardized storage of both the digital image data as well as other systematic meta-data collected about these images. The work resulting from my participation in several Pathology Informatics workshops has resulted in two publications this past year.
| Masuzzo, Paola; Martens, Lennart; 2014 Cell Migration Workshop Participants et al. (2015) An open data ecosystem for cell migration research. Trends Cell Biol 25:55-8 |
| Smith, Barry; Arabandi, Sivaram; Brochhausen, Mathias et al. (2015) Biomedical imaging ontologies: A survey and proposal for future work. J Pathol Inform 6:37 |
| Du, Wei; Cheung, Huey; Goldberg, Ilya et al. (2015) A Longitudinal Support Vector Regression for Prediction of ALS Score. IEEE Int Conf Bioinform Biomed Workshops 2015:1586-1590 |
| Eliceiri, Kevin W; Berthold, Michael R; Goldberg, Ilya G et al. (2012) Biological imaging software tools. Nat Methods 9:697-710 |
| Orlov, Nikita V; Eckley, D Mark; Shamir, Lior et al. (2012) Improving class separability using extended pixel planes: a comparative study. Mach Vis Appl 23:1047-1058 |
