Computer-Aided Diagnosis in Multiparametric MRI of the Prostate: An Open-Access Online Tool for Lesion Classification with High Accuracy

Abstract

Computer-aided diagnosis (CADx) approaches could help to objectify reporting on prostate mpMRI, but their use in many cases is hampered due to common-built algorithms that are not publicly available. The aim of this study was to develop an open-access CADx algorithm with high accuracy for classification of suspicious lesions in mpMRI of the prostate. This retrospective study was approved by the local ethics commission, with waiver of informed consent. A total of 124 patients with 195 reported lesions were included. All patients received mpMRI of the prostate between 2014 and 2017, and transrectal ultrasound (TRUS)-guided and targeted biopsy within a time period of 30 days. Histopathology of the biopsy cores served as a standard of reference. Acquired imaging parameters included the size of the lesion, signal intensity (T2w images), diffusion restriction, prostate volume, and several dynamic parameters along with the clinical parameters patient age and serum PSA level. Inter-reader agreement of the imaging parameters was assessed by calculating intraclass correlation coefficients. The dataset was stratified into a train set and test set (156 and 39 lesions in 100 and 24 patients, respectively). Using the above parameters, a CADx based on an Extreme Gradient Boosting algorithm was developed on the train set, and tested on the test set. Performance optimization was focused on maximizing the area under the Receiver Operating Characteristic curve (ROC_AUC). The algorithm was made publicly available on the internet. The CADx reached an ROC_AUC of 0.908 during training, and 0.913 during testing (p = 0.93). Additionally, established rule-in and rule-out criteria allowed classifying 35.8% of the malignant and 49.4% of the benign lesions with error rates of <2%. All imaging parameters featured excellent inter-reader agreement. This study presents an open-access CADx for classification of suspicious lesions in mpMRI of the prostate with high accuracy. Applying the provided rule-in and rule-out criteria might facilitate to further stratify the management of patients at risk.

Keywords: artificial intelligence; computer-aided diagnosis; machine learning; prostate cancer; prostate mpMRI.

Figure 1

Patient flow chart. An initial database search retrieved 142 patients eligible for this retrospective study. However, 18 patients were excluded based on the exclusion criteria. In total, 124 patients with 195 lesions were included. (*) Three lesions could not be unambiguously matched between the pathology report and the MRI assessment. The affected patients, however, featured other lesions that were clearly matched.

Figure 2

Boxplots of the acquired parameters. T2w signal intensity (SI), apparent diffusion coefficient (ADC), long and short diameter, dynamic contrast enhancement wash in and wash out, time to peak (TTP), peak enhancement intensity (PEI), initial area under the curve (iAUC), and arrival time (AT), and the additional parameters: prostate specific antigen (PSA) serum level, prostate volume, and patient age. Boxplots follow Tukey’s definition, with whiskers depicting the 1.5× interquartile ranges, and outliers marked as circles. The plots compare benign lesions (left boxes) with prostate cancer (PCa) lesions (right boxes), further divided into analyses of the whole gland (WG), and sub-analyses of the peripheral zone (PZ) and the central gland (CG). p-values, derived from Mann–Whitney U tests, are given above the plots.

Figure 3

Receiver Operating Characteristic curves of the CADx model determined from the 10-fold cross-validation (CV) procedure during training (black line; AUC = 0.908) and testing (grey line; AUC = 0.913), with no significant difference (p = 0.933).

Figure 4

Clinical case from the testing set. This 74-year-old patient was referred to radiology for prostate MRI because of an increase in serum PSA levels to 8 ng/mL and the suspicion of prostate cancer. Multiparametric MRI revealed a 10 × 7 mm PI-RADS 4 lesion dorsally in the mid/apical peripheral zone (green arrow; classification according to PI-RADS v2.1 [6]). The lesion featured high signal intensity in DWI (a), along with a strong diffusion restriction in ADC (b). T2w axial images showed a signal decrease in the sharply delineated lesion (c). The lesion featured a steep wash-in curve, along with a decent wash out (d,e). Subsequent histopathology revealed scarred prostate tissue and few prostate glands with urothelial metaplasia, but no evidence of malignancy. The CADx correctly excluded malignancy for this peripheral zone PI-RADS 4 lesion, with a false negative rate of 3%. The patient received a follow-up examination 9 months later in which the lesion had remained unchanged.

Figure 5

Interface of the open-access web application, accessible at http://bit.do/Prostate-MRI, with an analysis of the peripheral zone PI-RADS 4 lesion presented in Figure 4. The lesion was correctly classified as benign, with an error rate of 3%. Its location on the ROC curve is highlighted with a red circle. Sensitivity: 97.0% (95% CI: 89.6–99.6%); NPV: 96.2% (95% CI: 87.0–99.5%); negative likelihood ratio: 0.052 (95% CI: 0.013–0.206).

Figure 6

Image assessment. (a,b) Prostate volume was calculated from the diameters measured in axial and sagittal T2w images, as described in the PI-RADS v2.1 guidelines [6]. A circular region of interest (ROI) was placed within a lesion, for lesions of the peripheral zone, preferably on the ADC map (c) in additional consideration of DWI (calculated b-value of 1500 s/mm²; (d)), and for lesions of the transitional zone/central gland, preferably in the axial T2w sequence (e) [5,6]. Following placement, the ROI was copied to the other sequences using the software. If lesion measurement was difficult or compromised on ADC (for peripheral zone lesions) or T2w (for transitional zone/central gland lesions), measurements were performed on the sequence that showed the lesion best [5,6], e.g., in the above example, lesion dimensions were assessed on DWI (d). (f,g) Time-intensity curves were automatically calculated by the software, yielding the dynamic parameters: wash in, wash out, time to peak (TTP), peak enhancement intensity (PEI), arrival time (AT), and initial area under the curve (iAUC).