Evaluation of machine learning techniques for prostate cancer diagnosis and Gleason grading

Although the gold standard for prostate cancer tissue grading has been the Gleason grading scheme, it is strongly affected by 'inter- and intra observer variations'. Therefore, the development of objective and reproducible computer-aided classification methods is of critical importance. In this paper, 16 supervised machine learning algorithms were compared based on their performance on prostate cancer diagnosis and Gleason grading. The classification problems addressed were: tumour vs. non-tumour, low vs. high grade; and the four class problem of diagnosis and grading. Thirteen Haralick texture characteristics were calculated based on grey level co-occurrence matrix of microscopic prostate tissue. For the best performing algorithm in each case the accuracy obtained was 97.9% for diagnosis (tumour-non-tumour), 80.8% for low-high grade discrimination and 77.8% for accomplishing both diagnosis and Gleason grading. Logistic regression and sequential minimal optimisation for training a support vector machine were among the four top scoring algorithms in each classification problem.