Classification of suspicious lesions on prostate multiparametric MRI using machine learning

Prostate cancer is the most common cancer in American men.1 Clinical decisions, related to the need for prostate biopsy, which areas to biopsy, and regions that require attention in focally directed therapy, are multifactorial and complex. The use of multiparametric MRI (mpMRI) for detecting prostate cancer is rapidly evolving because of its growing availability and ability to combine functional [perfusion via dynamic contrast-enhanced (DCE-MRI) and diffusion via diffusion-weighted imaging (DWI)], and anatomical information [T2-weighted (T2-W) MRI]. The combined information in the multiple mpMRI sequences results in high sensitivity and specificity for distinguishing intraprostatic cancer of higher grade [Gleason score (GS) 7 or above].2,3 To standardize the evaluation and reporting of prostate mpMRI, an expert panel of the European Society of Urogenital Radiology published in 2012 the prostate imaging reporting and data system (PI-RADS) for prostate cancer detection4 (current version PI-RADSv.2).5 Unlike PI-RADS, which largely depends on subjective assessment of mpMRI, computer-aided diagnosis (CAD) techniques for quantitative mpMRI analysis have also been developed for prostate cancer detection and diagnosis.6–12 The core components in CAD systems for prostate cancer, as summarized by Liu et al.13 and Lemaitre et al.,14 include preprocessing of the images, segmentation of the prostate, image registration between MRI modalities, feature extraction, and voxel classification. The CAD efforts can be divided into two categories based on the main objectives for the analysis: (i) detection/segmentation of the suspicious lesion and/or (ii) assessment of the aggressiveness of prostate cancer.15 In 2016, the International Society for Optics and Photonics (SPIE), along with the support of the American Association of Physicists in Medicine (AAPM) and the National Cancer Institute (NCI), conducted a “grand challenge” on quantitative image analysis methods for the diagnostic classification of clinically significant prostate lesions.16 As opposed to the detection task, the location of the suspicious lesion in the SPIE-AAPM-NCI PROSTATEx challenge was provided and the goal was to decide if the lesion is related to clinically significant cancer (GS ≥ 7) or benign and/or low risk (GS = 6). The PROSTATEx challenge task was different from determining whether a lesion is cancerous or not,7,9,17,18 or discriminating between GS (3 + 4) = 7 and GS (4 + 3) = 7.19–21 However, the general approach for these feature-based algorithms is to compute a set of quantitative imaging features from the mpMRI data and develop a supervised classifier using the computed features from the training cases and their associated “ground-truth” labels. The developed classifier is then used to classify new cases (test dataset). The increased applications of prostate mpMRI for clinical decision making, and the plethora of the associated imaging sequences and quantifiable features, bring forth the following questions: What are the optimal ways to analyze mpMRI data for the prostate; do all MR sequences contribute to the accuracy of the MRI interpretation; which of the imaging features have importance for the final classification? In this paper, we present our approach for analysis of the PROSTATEx challenge data to assess the aggressiveness of the suspicious lesions.16 A collection of retrospective set of prostate MR studies was provided to the challenge participants. The data were separated in training and test datasets. The lesions in the training dataset were labeled with “TRUE” and “FALSE” for the presence of significant cancer. The assignment of the lesions in the test dataset was held back by the challenge organizers. The goal was to develop a model, based on the training dataset and using this model to classify the lesions in the test dataset. The selection of the prediction models was based on the area under the ROC curve (AUC). The predicted scores (0 to 1, 0 = benign, 1 = significant cancer) for each lesion in the test dataset were obtained using the prediction models. The predicted scores were sent to the challenge organizers and the models evaluated by AUC. The AUC results for the test dataset were communicated back to the challenge participants. The PROSTATEx challenge provided us with a unique opportunity to compare our algorithms with those of others from academia, industry, and government. We evaluated all sequences of mpMRI and investigated the importance of the imaging features to the final classification model.