Proof-of-Concept and Performance Studies

Two of the lead institutions for the PathMiner project (UMDNJ and UPenn) previously collaborated on a project involving next generation imaging and telecommunications technologies.

Through grant funding from the federal Defense Advanced Research Projects Agency (DARPA), UMDNJ and Telcordia Technologies (formerly BellCore) partnered with several other leading institutions to establish a regional consortium for evaluating the use of Virtual Private Networks for large-scale research applications in telemedicine and medical imaging. UMDNJ's Image Guided Decision Support (IGDS) prototype and Distributed Telemicroscopy (DT) prototype were two of the key technologies that were utilized in performance studies. In June, 2002 Dr. Foran's laboratory worked with investigators from Telcordia, OSU, and Pitt to host a site visit for the Program Manager at DARPA. After making formal presentations to the panel the team conducted a real-time, Internet-based demonstration of the IGDS and the DT as they ran across a tri-state area using AA-VPN communications and conditions of network contention. Since that time the PathMiner project has been presented regularly at the annual meeting of the national Association of Pathology Informatics, during a recent invited presentation at the Federation of American Societies for Experimental Biology (FASEB) meeting in Washington, DC and was recently featured on the Internet2 website.

A snapshot of the Distributed Telemicroscopy (DT) and Image Guided Decision Support client interfaces are shown in Figures 1 and 2, respectively.

The Image Guided Diagnosis Support (IGDS) system provides the means for multi-modal indexing, querying, and retrieval of pathology specimens and correlated clinical data based upon the visual content of the constituent images. Correspondingly, the system also enables one to submit a specific protein or molecular profile and retrieve a ranked set of digitized cells exhibiting that signature. Using advanced computer vision techniques the system was recently modified to support unsupervised analysis in order to deliver compositional results for a region of interest and can be configured to query multiple distributed databases across the Internet.

Based upon the improvements that were realized in performing classifications and the progress that was made in performing completely unsupervised scanning and imaging of specimens our team began to conduct experiments to explore the use of the image-guided approach for a wider range of disorders which extended beyond the scope of the original proposal by including acute myeloid leukemia (AML) and acute lymphocytic leukemia (ALL).

These studies were conducted using a mixed set of nearly 4,000 imaged cells taken from 105 different patients originating from four different hospitals (City of Hope National Medical Center, Duarte, CA; Hospital of the University of Pennsylvania, Philadelphia, PA; Robert Wood Johnson University Hospital, NJ; and Spectrum Health System, Grand Rapids, MI). There were obvious variations in the condition and staining characteristics of these specimens because of differences in manufacturers of dyes, choices in automated staining devices, and due to the overall intensity values. All of these variables led to variations in shadowing, shading, contrasts, and highlighting cues which made this task especially challenging, but also provided detailed insight as to the performance in real-case, multi-institutional scenarios. As a natural extension of our prior work we maintained our focus on utilizing texture descriptors, but invested considerable effort in performing a systematic investigation of both the feature measures used and the methods implemented to perform the classification.

Our team already performed a systematic investigation to determine the optimal number of modes for differentiating among these disorders. Through the use of texton histograms, aggregated classifiers and Support Vector Machines (SVMs) we achieved nearly 85% correct classification performance on a cell by cell basis and 89% correct classification for a case by case basis in spite of the increased number of classes and the large variation in staining characteristics exhibited by these specimens.