Validation of Automated Image Analysis for Hematopathology
Bryan Dangott, Nisha Ramesh, Tolga Tasdizen, Mohamed Salama. University of Utah, Salt Lake City, UT
Background: Whole slide image (WSI) analysis has great potential for standardizing diagnostic interpretation, streamlining workflow, and improving patient care. Automated differential counting of whole blood smears is one potential challenging application due to the size of the WSI files, inherent image artifact, and stain variability. There are also limitations contingent on properly selecting the relevant areas of interest. Our goal was to evaluate the effectiveness of an automated region of interest (ROI) algorithm by comparing it to manual ROI methods.
Design: Five randomly selected peripheral blood smears were scanned at 20x using an Aperio CS scanner. In the manual group (MG), the areas to be analyzed were selected by a trained user that visually selected high power snapshots from WSI of peripheral smears. In the automated group (AG), Definiens Developer XD was used to automatically detect the ROI and extract the white cells. MG and AG data served as independent data sets which were analyzed by a standardized MATLAB application. The application subclassified the white cells using linear discriminant analysis. All of the white cells were accurately classified by a hematopathologist to measure any differences in performance of the MATLAB application between the groups. The MATLAB application served as an objective benchmark of the inputs (MG and AG) with respect to the desired, clinically useful output (classification).
Results: 1919 cells were evaluated in the MG. 3190 cells were evaluated in the AG. The MATLAB classifier was most successful with the neutrophil cell line (98% accuracy for both MG and AG). The eosinophils had the lowest classification accuracy (67% for MG, and 72% for AG). There were no significant differences in percent classification between MG and AG for any of the cell types. Thus the algorithm to automatically select the ROI performed as well as a trained user who manually selected the ROI.
Conclusions: This novel technique using WSI to perform differential counts via automated techniques allows streamlining of workflow. The labor intensive portions of the differential count can be automated with subsequent review and fine tuning by the pathologist. Excellent concordance was achieved between manually selected ROI inputs and automatic ROI techniques.
Tuesday, March 20, 2012 1:00 PM
Poster Session IV # 221, Tuesday Afternoon