Characterization of Prostate Cancer aggressiveness based on bi-parametric MRI.

The aim of this study is to provide a noninvasive, radiological image-based Computer Aided Diagnosis (CAD) able to distinguish between high aggressive (Gleason Score (GS) >= 4+3) and low-aggressive (GS<= 3+4) Prostate Cancers (PCa). The system exploits the use of Machine Learning (ML) and Deep Learning (DL) on biparametric Magnetic Resonance Images coming from Candiolo IRCCS and San Giovanni Molinette hospital. Regarding the ML approach, once tumor areas have been manually segmented, features of first order statistics, intensity-based and texture features are extracted, both from T2WI and ADC maps. The study carries out a parallel analysis of ten different Datasets, which differ in type of feature (3D or 2D), voxel spacing, application of filters, and bin number. Datasets have been pre-processed using data cleaning techniques, then Univariate Analysis (UA) and Multivariable Analysis (MA) are carried out. The UA involves the calculation of the area under the ROC curves (AUC) of each feature, Mann Witney U test and correlation analysis, both between each feature vector and the output (classification). The MA includes the Genetic Algorithm (GA), the Minimum Redundancy Maximum Relevance (MRMR) and the Affinity Propagation (AP) methods. SVM classifiers have been optimized, using four Feature Selection strategies. The first one consists of evaluating the 7-fold cross-validation performances of the model trained with an increasing number of features, added one by one in descending order of AUC, until the overfitting point is found. The others use the subsets resulting from the three multivariable algorithms. At the end, the best ML classifier is a T2 SVM model with polynomial kernel, trained with features selected by GA. It obtained a 100% accuracy in the Training Set (TRS), high performance in terms of accuracy (93.75%), specificity (83,33%) and sensitivity (100%) in the Test Set (TSS), but these decreased on the validation set (62.07%, 70.59%, and 50% respectively). Regarding the DL approach, once the ROIs (3x3 and 5x5 pixel, totally inside the lesion) have been extracted, both from T2WI and ADC maps, Convolutional Neural Networks (CNN) with 1, 2, and 3 Convolutional Layers are tested. Several CNNs are trained, different in size and number of filters, number of neurons, and set parameters. For the dataset division, we proceeded at first, as in the ML part, maintaining the Molinette lesions as an external validation set and dividing the ROIs randomly into TRS and TSS, and, then, selecting the lesions at random, so that ROIs belonging to the same lesion could not be present in both the TRS and TSS and adding some Molinette lesions in the TRS. The resulting best DL classifier is a model trained on T2 ROI 5x5 with 3 convo- lutional layers. The performances obtained in terms of accuracy, specificity, and sensitivity on the MRI slices are: 96.2%, 96.3% and 96.1% in the TRS; 62.5%, 87.5% and 54.2% in the TSS; 44.9%, 46.2% and 43.5% in the validation set. The results from ML and DL approaches show lower results in the validation sets due to the low ability of the classifiers to generalize the problem. In particular, the best ML model achieves better performance than the best DL one, even if the latter one has been trained with a bigger TRS. The generalization problem must be reduced increasing the number of samples in the datasets, this could increase the performance of both ML and DL models.