2006 Scientific Session Abstracts
E-Poster Presentations
Submission Compliance and Guidelines
Copyright Transfer Form (PDF)
Abstract Archive
A Hierarchical Computer-aided Classification Scheme for Automated Detection of Prostatic Adenocarcinoma from Digitized Histology
Michael Feldman, MD, PhD1, (feldmanm@mail.med.upenn.edu); John Tomaszeweski, MD, PhD1; 1Department of Surgical Pathology, University of Pennsylvania, Philadelphia, Pennsylvania; Scott Doyle2, Anant Madabhushi, PhD2 ; 2 Department of Biomedical Engineering, Rutgers University, Piscataway, New Jersey
Content: We present a computer-aided detection (CAD) system that integrates several hundred image features via a hierarchical classifier to automatically distinguish between benign and malignant tissue regions on digitized histological images of prostate biopsy samples. The images are first analyzed at the lower coarser resolutions or scales and the initial approximate cancer segmentations are then analyzed in greater detail at the higher image scales.
Technology: A set of H&E stained prostate tissue samples were scanned into a computer using a high-resolution glass tissue slide scanner and saved as TIFF images. These images were decomposed using a standard pyramidal scheme into the constituent image scales. All image analysis routines were implemented using the MATLABTM software package.
Design: The CAD scheme extracts over 600 image texture features at the lowest image scale. Bayesian classification is used to assign each image pixel a likelihood of malignancy based on the value of each feature. The AdaBoost algorithm is used to select only the most discriminatory features which are then combined to produce an initial coarse segmentation of cancer. The process is repeated at the subsequent higher image scales to successively refine the segmentation. Since the computations at the higher image scales are limited to only those regions determined as cancer at the preceding lower scales, the scheme is highly efficient and accurate.
Results: The CAD scheme was found to have an average accuracy of 94% in distinguishing between benign and malignant regions on 20 patient studies. Further, as can be observed from the ROC curves in Figure 1, while the cancer detection sensitivity of the CAD system remains relatively constant, the specificity significantly increases at the higher scales.
Figure 1: ROC curves showing the increase in cancer detection specificity from the lower (solid red line) to the higher scales (dashed lines). Note that the detection sensitivity remains relatively constant across scales.
Conclusion: We have presented a highly accurate and efficient CAD system for detecting prostate cancer from digitized histology. The use of a hierarchical classification paradigm results in a nearly 8 fold savings in computational time. Evaluation of CAD against manually determined cancer ground truth reveals an average classification accuracy of 94%. These preliminary results suggest that the system has the potential to become a diagnostic tool to assist pathologists.