Duc Fehr PhD, Presenter: Nothing to Disclose

Harini Veeraraghavan, Abstract Co-Author: Nothing to Disclose

Andreas Georg Wibmer MD, Abstract Co-Author: Nothing to Disclose

Hebert Alberto Vargas MD, Abstract Co-Author: Nothing to Disclose

Evis Sala MD, PhD, Abstract Co-Author: Nothing to Disclose

Hedvig Hricak MD, PhD, Abstract Co-Author: Nothing to Disclose

To develop a machine learning-based automatic feature selection for classification of PCa and the associated Gleason Score (GS) from multiparametric prostate MRI (mpMRI).

158 prostate cancer patients who underwent mpMRI within 6 months prior to prostatectomy were retrospectively analyzed. Volumes of interest were placed in cancerous and normal peripheral zone on T2-weighted MRI (T2WI) and apparent diffusion coefficient (ADC) maps, using step-section pathology maps of the surgical specimens as reference. Statistical image features (mean, standard deviation, skewness, kurtosis) and Haralick texture features (energy, entropy, correlation, homogeneity, contrast) were computed from these maps. Adaptive Boosting using support vector machine (AdaBoost-SVM) machine learning was applied to extract salient features and learn the best classification model. Robust classifier performance was obtained through 10-fold crossvalidation. In each fold a small percentage of the samples was kept for testing, while the rest was used for training. Thus, the testing was done with novel data whose true classification labels were unknown to the classifier.

The algorithm achieved an accuracy of 93% for classifying cancerous vs normal structures and 83% for classifying GS (6/7+). The algorithm extracted ADCentropy, T2kurtosis, T2mean, and ADCenergy as features for cancer vs normal tissue and ADCkurtosis, T2entropy, T2correlation, and ADCcontrast for GS classification. A statistical t-test analysis confirms the salient features found by our approach for normal vs cancerous tissue: ADCentropy (p<0.001), T2kurtosis (p<0.001), T2mean (p=0.45), ADCenergy (p<0.001). For GS classification, T2entropy (p=0.03) was significant.  

We developed an algorithm that extracts salient features from MRI and classifies PCa and GS. The relevance of machine learning extracted features was confirmed by t-test. The extracted features can be used to generate new images that can potentially assist radiologist interpretation.  

Image-based automatic prostate cancer and GS classification can assist radiologists in interpreting MRI and contribute to patient risk-stratification and treatment selection.  

Fehr, D, Veeraraghavan, H, Wibmer, A, Vargas, H, Sala, E, Hricak, H, Automatic Classification of Prostate Cancer and Gleason Scores through Machine Learning and Salient Feature Selection from Multiparametric MRI.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14014774.html