The computer-aided analysis of prostate multiparametric MRI (mpMRI) could improve significant-prostate-cancer (PCa) detection. Various deep-learning-and radiomics-based methods for significant-PCa segmentation or classification have been reported in the literature. To be able to assess the generalizability of the performance of these methods, using various external data sets is crucial. While both deep-learning and radiomics approaches have been compared based on the same data set of one center, the comparison of the performances of both approaches on various data sets from different centers and different scanners is lacking. The goal of this study was to compare the performance of a deep-learning model with the performance of a radiomics model for the significant-PCa diagnosis of the cohorts of various patients. We included the data from two consecutive patient cohorts from our own center (n = 371 patients), and two external sets of which one was a publicly available patient cohort (n = 195 patients) and the other contained data from patients from two hospitals (n = 79 patients). Using multiparametric MRI (mpMRI), the radiologist tumor delineations and pathology reports were collected for all patients. During training, one of our patient cohorts (n = 271 patients) was used for both the deep-learning-and radiomics-model development, and the three remaining cohorts (n = 374 patients) were kept as unseen test sets. The performances of the models were assessed in terms of their area under the receiver-operating-characteristic curve (AUC). Whereas the internal cross-validation showed a higher AUC for the deep-learning approach, the radiomics model obtained AUCs of 0.88, 0.91 and 0.65 on the independent test sets compared to AUCs of 0.70, 0.73 and 0.44 for the deep-learning model. Our radiomics model that was based on delineated regions resulted in a more accurate tool for significant-PCa classification in the three unseen test sets when compared to a fully automated deep-learning model.

Radiomics applied in MRI has shown promising results in classifying prostate cancer lesions. However, many papers describe single-center studies without external validation. The issues of using radiomics models on unseen data have not yet been sufficiently addressed. The aim of this study is to evaluate the generalizability of radiomics models for prostate cancer classification and to compare the performance of these models to the performance of radiologists. Multiparametric MRI, photographs and histology of radical prostatectomy specimens, and pathology reports of 107 patients were obtained from three healthcare centers in the Netherlands. By spatially correlating the MRI with histology, 204 lesions were identified. For each lesion, radiomics features were extracted from the MRI data. Radiomics models for discriminating high-grade (Gleason score ≥ 7) versus low-grade lesions were automatically generated using open-source machine learning software. The performance was tested both in a single-center setting through cross-validation and in a multi-center setting using the two unseen datasets as external validation. For comparison with clinical practice, a multi-center classifier was tested and compared with the Prostate Imaging Reporting and Data System version 2 (PIRADS v2) scoring performed by two expert radiologists. The three single-center models obtained a mean AUC of 0.75, which decreased to 0.54 when the model was applied to the external data, the radiologists obtained a mean AUC of 0.46. In the multi-center setting, the radiomics model obtained a mean AUC of 0.75 while the radiologists obtained a mean AUC of 0.47 on the same subset. While radiomics models have a decent performance when tested on data from the same center(s), they may show a significant drop in performance when applied to external data. On a multi-center dataset our radiomics model outperformed the radiologists, and thus, may represent a more accurate alternative for malignancy prediction.

Significant prostate carcinoma (sPCa) classification based on MRI using radiomics or deep learning approaches has gained much interest, due to the potential application in assisting in clinical decision-making. Objective: To systematically review the literature (i) to determine which algorithms are most frequently used for sPCa classification, (ii) to investigate whether there exists a relation between the performance and the method or the MRI sequences used, (iii) to assess what study design factors affect the performance on sPCa classification, and (iv) to research whether performance had been evaluated in a clinical setting Methods: The databases Embase and Ovid MEDLINE were searched for studies describing machine learning or deep learning classification methods discriminating between significant and nonsignificant PCa on multiparametric MRI that performed a valid validation procedure. Quality was assessed by the modified radiomics quality score. We computed the median area under the receiver operating curve (AUC) from overall methods and the interquartile range. Results: From 2846 potentially relevant publications, 27 were included. The most frequent algorithms used in the literature for PCa classification are logistic regression (22%) and convolutional neural networks (CNNs) (22%). The median AUC was 0.79 (interquartile range: 0.77–0.87). No significant effect of number of included patients, image sequences, or reference standard on the reported performance was found. Three studies described an external validation and none of the papers described a validation in a prospective clinical trial. Conclusions: To unlock the promising potential of machine and deep learning approaches, validation studies and clinical prospective studies should be performed with an established protocol to assess the added value in decision-making.