Jeremy Bancroft Brown, Presenter: Nothing to Disclose

Maryellen L. Giger PhD, Abstract Co-Author: Stockholder, Hologic, Inc Royalties, Hologic, Inc Royalties, General Electric Company Royalties, MEDIAN Technologies Royalties, Riverain Medical Royalties, Mitsubishi Corporation Royalties, Toshiba Corporation

Li Lan MS, Abstract Co-Author: Nothing to Disclose

Robert Tomek BS, Abstract Co-Author: Employee, Quantitative Insights, Inc

Variable assessment of prostate cancer by radiologists and pathologists limits the reliability of the truth used to train CAD algorithms. We investigated the effect of differing radiologist / pathologist assessments upon computerized texture analysis on T2-weighted MR images, and the effect in the development of MRI-based CAD algorithms for prostate cancer.

Prostatectomy specimens from 54 patients were reviwed in consensus by a radiologist / pathologist pair (pair A). The radiologist delineated cancer foci on T2-weighted prostate MR images, and the pathologist determined Gleason scores. Forty-three of these specimens were then reviewed in the same manner by a second radiologist / pathologist pair (pair B). Texture features based upon a gray level co-occurrence matrix (GLCM) were extracted from each MR lesion, and two were selected for further analysis on the basis of their ROC analysis AUC for separating GS = {6,7} cases from GS={8,9} cases, as well as their low correlation with the other texture features. These two texture features were independently analyzed using a two-way fixed effect ANOVA model with interaction. The model included the effect of Gleason score (6, 7, 8, or 9) and radiologist / pathologist pair (A or B) on the texture features.

Of the 55 (pair A) lesions and 64 (pair B) lesions, twenty-nine were in nearly the same location between the two pairs of physicians, while 61 were distinct. Of the corresponding lesions, the two pathologists assigned the same Gleason score to 21. For the ANOVA on texture features, the effect of Gleason score upon the variation of feature 1 was significant (F = 3.38; p = 0.02), while the effects of the radiologist / pathologist pair and the interaction term were not (p = 0.88 and p = 0.89, respectively). The effect of Gleason score upon the variation of feature 2 was close to significant (F = 2.07; p = 0.11), while the effects of the radiologist / pathologist pair and the interaction term were not (p = 0.97 and p = 0.50, respectively).

Gleason score (as opposed to the specific pathologist determining the score or radiologist delineating the cancer) is the dominant effect upon GLCM-based texture features on T2-weighted prostate MRI.

This study indicates that a clinical MRI-based CAD system for prostate cancer diagnosis should be feasible, despite the variability of the truth that would be used to train such a system.

Bancroft Brown, J, Giger, M, Lan, L, Tomek, R, Variability in Radiologist/Pathologist Assessment of Prostate Cancer: Implications for Computerized Analysis of T2-weighted MR Images.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11034555.html