Pergamon 0301-5629(94)00094-8 Ultrasound in Med. & Biol., Vol. 20, No. 9, pp. 871-876, 1994 Copyright © 1994 Elsevier Science Ltd Printed in the USA. All rights reserved 0301-5629/94 $6.00 + .00 OOriginal Contribution THE RELIABILITY OF COMPUTER ANALYSIS OF ULTRASONOGRAPHIC PROSTATE IMAGES: THE INFLUENCE OF INCONSISTENT HISTOPATHOLOGY R. J. B. GIESEN, A. L. HUYNEN, J. J. M. C. H. DE LA ROSETTE, H. E. SCHAAFSMA, M. P. VAN IERSEL, R. G. AARNINK, F. M. J. DEBRUYNE and H. WIJKSTRA Department of Urology, University Hospital Nijmegen, The Netherlands (Received 10 January 1994; in final form 18 May 1994) Abstract--This article describes a method to investigate the influence of inconsistent histopathology during the development of tissue discrimination algorithms. Review of the pathology is performed on the biopsies used as training set of a computer system for cancer detection in ultrasonographic prostate images. The influence of the discrepancies found between independent pathologists on the discriminating power of the system is investigated. A high diagnostic consistency in histopathology concerning only the categories malignant and nonmalignant is found. Therefore, review of the pathology does not significantly influence the results of tissue discrimination algorithms for cancer detection. However a high interobserver variability is obtained in the differentiation between more histology classes. Key Words: Histopathology, Tissue characterization, Ultrasonography, Prostate. INTRODUCTION In the clinical environment histopathology is used as "gold standard" for tissue discrimination. Therefore, during the development and evaluation of algorithms for tissue characterisation, histopathology is often used as reference. Consequently, the reliability of the devel- oped methods and corresponding results are strongly dependent on the reliability of histopathology. Never- theless, pathology is sometimes ambiguous, and cate- gorization and grade have limitations. Robertson et al. (1989) presented a good agreement between 12 consul- tant histopathologists in reporting of CIN-3 and squa- mous carcinoma in cervical biopsy specimens, but an inability to distinguish between the lesser grades of CIN. Also a high interobserver variability in the identi- fication of cellular changes associated with infection was reported. This group also presented the results of examination of 90 urinary bladder biopsies by 11 pathologists (Robertson et al. 1990). This study showed reasonable agreement in the grading and stag- ing of transitional cell carcinoma and in the diagnosis of high grade dysplasia, but very poor agreement for lesser degrees of hyperplasia was reported. Beck (1985) described 82% agreement between the diagno- Address correspondence to: R. J. B. Giesen, Department of Urology, University Hospital Nijmegen, P.O. Box 9101, 6500 HB, Nijmegen, The Netherlands. 871 sis of breast lesions. It was also shown that each pathol- ogist was generally consistent in either under- or over- diagnosing. Structural differences between different pathologists were found. At our department, histopathology is used as gold standard for automated tissue discrimination in ultraso- nographic prostate images. A system for quantification of textural features in ultrasonographic images of the prostate has been developed; the Automated Urologic Diagnostic EXpert (AUDEX) system (Giesen and Huy- hen 199l; Huynen et al. 1994). Although ultrasound is a widely used imaging tool for visualization of the human prostate (Waterhouse and Resnick 1989; Lee et al. 1989; Shinohara et al. 1989), the interpretation of transrectal ultrasonographic prostate images is sub- jective and highly dependent on the experience and expertise of the urologist (Bertermann et al. 1989; Loch et al. 1990; Scardino et al. 1989; Shinohara et al. 1989). The AUDEX system supports the urologist in his inter- pretation of ultrasonographic prostate images. Addi- tional information is provided by colour coding these images according to the probability of malignancy of the tissue visible in the image. To provide discrimina- tion between benign and malignant tissue, the system was trained with images from tissue biopsies. Using these images, a correlation between texture-describing parameters calculated at the puncture location and his-