Multi-scale lacunarity as an alternative to quantify and diagnose the behavior of prostate cancer

Prostate cancer is a serious public health problem accounting for up to 30% of clinical tumors in men. The diagnosis of this disease is made with clinical, laboratorial and radiological exams, which may indicate the need for transrectal biopsy. Prostate biopsies are discerningly evaluated by pathologists in an attempt to determine the most appropriate conduct. This paper presents a set of techniques for identifying and quantifying regions of interest in prostatic images. Analyses were performed using multi-scale lacunarity and distinct classification methods: decision tree, support vector machine and polynomial classifier. The performance evaluation measures were based on area under the receiver operating characteristic curve (AUC). The most appropriate region for distinguishing the different tissues (normal, hyperplastic and neoplasic) was defined: the corresponding lacunarity values and a rule's model were obtained considering combinations commonly explored by specialists in clinical practice. The best discriminative values (AUC) were 0.906, 0.891 and 0.859 between neoplasic versus normal, neoplasic versus hyperplastic and hyperplastic versus normal groups, respectively. The proposed protocol offers the advantage of making the findings comprehensible to pathologists. (C) 2014 Elsevier Ltd. All rights reserved.