Fig 1.
Flow of participants though the study.
STARD diagram to demonstrate the flow of participants though the study. Downstream of the “Index test” the analysis is lesion-based due to detection of multiple lesions in some patients. PSA, Prostate Specific Antigen; CAD, Computer-Aided Detection; TP, True Positive; TN; True Negative; FP, False Positive; FN, False Negative.
Table 1.
Patient demographics and database description.
Table 2.
Sequences used for multiparametric magnetic resonance imaging.
Fig 2.
Sequences used for multiparametric MRI (mpMRI) imaging according to PI-RADS™.
Sample images from lesions (white ROIs in A-C) histologically characterized as (A) prostatitis (B) hyperplasia (BPH) and (C) adenocarcinoma (PCa). (i) T2w TSE images, (ii) apparent diffusion coefficient (ADC) tables based on DWI b0-100-500-800-1000 s/mm2 (iii) T1w FFE with enhancement and (iv) sample Dynamic Contrast Enhanced (DCE) curves of the lesion core, defined as red colored ROI in the T1w FFE images. Scale bar 3cm. For MRI sequence parameter details see Table 2.
Fig 3.
Automated mpMRI classification and histological ground truth.
Target lesions were manually drawn by human observers. Sample images for (A-C) classified and (D, E) non classified PCa and benign biopsy cores. (A) Prostatitis, (B) adenocarcinoma (PCa) Gleason grade 6, (C) PCa Gleason grade 9, (D) atypical small acinar proliferation (ASAP) and (E) PCa Gleason grade 7. From left to right: (i) T2w with MAI-heatmap and outlined lesions (white line), (ii) Hematoxylin-Eosin histopathology of the corresponding biopsy cores and (iii) MAI histograms. Warm colors in MAI heatmaps (i) represent higher values in a scale 0–1. Classified lesions revealed a “warm-colored” MAI-map and a left-skewed histogram.
Fig 4.
Classification of malignant and benign biopsy cores.
(A) Proportion of classified to non classified malignant and benign lesions. From 104 mixed type cores 36 were classified, from those (classified/total) 14/57 were benign and 22/47 malignant. (B) The classification was independent of the Gleason grade, P 0.713 R 0.193 Pearson correlation (C) Classification of benign lesions. ASAP, Atypical Small Acinar Proliferation; FMD, Fibromuscular Dysplasia; BPH, Benign Prostatic Hyperplasia.
Table 3.
Identity of classified and non classified lesions.
Table 4.
CAD sensitivity and specificity for PCa detection.
Fig 5.
MAI as malignancy biomarker and MAI correlation with the Gleason grade.
(A) Average ADC values (x10-6 mm2/s) of observed lesions, 650/295 (median/IQR) for malignant and 950/339.5 for benign, N (malignant/benign) 47/57, P < 0.0001 Mann-Whitney U-test (B) MAI of observed lesions, 0.56/0.23 (median/IQR) for malignant and 0.39/0.18 for benign lesions, N (malignant/benign) 47/57, P 0.023 Mann-Whitney U-test. (C) MAI in benign, malignant, classified and non classified biopsy cores, P < 0.001 Kruskal-Wallis ANOVA on ranks, * P < 0.05 Dunn´s post hoc test (D) No significant correlation between MAI and Gleason grade, P 0.522, N 22, R 0.144, Pearson´s product moment correlation; n.s., non significant.
Fig 6.
Receiver operating characteristic (ROC) trade-off curve for ADC, MAI and PI-RADS.
ADC mean (black), MAI median (dark green), MAI mean (light blue) and MAI median/mean ratio (green). MAI is compared with the ADC performance alone and with the observer´s performance according to PI-RADS™ v1 (dark blue) and PI-RADS™ v2 (violet). The area under the curve (AUC) is 0.64±0.057 (mean, SEM) with 95% CI 0.53–0.75 and P 0.02 for MAI mean, 0.63±0.058 with 95% CI 0.52–0.74 and P 0.02 for MAI median and 0.59±0.058 with 95% CI 0.47–0.70 and P 0.13 for the MAI median/mean ratio. Corresponding values for the mean ADC lesion value are AUC 0.79±0.05 with 95% CI 0.70–0.88, P < 0.0001. Observer´s performance for PI-RADS™ v1 was AUC 0.67±0.05 with CI 0.58–0.76, P 0.003 and for PI-RADS™ v2 AUC 0.68±0.04 with CI 0.59–0.76, P 0.002. N malignant/benign cores 47/57. MAI and PI-RADS (v1, v2) reveal comparable performances in malignancy detection, P 0.60 for MAI vs PI-RADS v1 and P 0.53 for MAI vs PI-RADS v2, chi-squared test. ADC is superior to MAI in malignancy prediction, P 0.04, chi-squared test.
Table 5.
Receiver operating characteristic (ROC) analysis for MAI and PI-RADS.
Fig 7.
CAD performance is reliable for lesions larger than 1 ml.
(A) Counts of classified and non classified cores in relation to their volume. Volume distribution of all lesions (number, %): 0–0.5ml (51, 49.04), 0.5–1.0ml (38, 36.54) and larger than 1.0ml (15, 14.42). True positive (TP, blue) lesions include classified PCa, true negatives (TN, cyan) are the non classified benign cores, false positives (FP, yellow) are the classified benign cores and false negatives (FN, red) the non classified PCa. The CAD-sensitivity increases and the number of FN decreases towards larger lesion volumes: (sensitivity % / FNR %) 27.27/31.37 for 0–0.5ml lesions, 53.33/18.42 for 0.5–1.0ml lesions and 80.00/13.33% for lesions larger than 1.0ml. (B) MAI score with lesion volume. A strong trend for a positive correlation between lesion size and MAI-score was found for TP lesions (P 0.057, Pearson´s correlation) but not for any of the remaining categories (TN, FP and FN, P > 0.1, Pearson´s correlation). (C) Lesions smaller than 0.5 ml show the same malignancy incidence and comparable aggressiveness compared to larger lesions (Ci) Lesions smaller than 0.5 ml (number, %) benign (29, 56.86) malignant (22, 43.14), (Cii) Gleason histogram for malignant lesions smaller than 0.5 ml, (Ciii) Lesions larger than 0.5 ml (number, %) benign (28, 57.14) malignant (21, 42.86), (Civ) Gleason histogram for malignant lesions larger than 0.5 ml.