Detecting Clinically Significant Prostate Cancer in PI-RADS 3 Lesions Using T2w-Derived Radiomics Feature Maps in 3T Prostate MRI

Abstract

Prostate Imaging Reporting and Data System version 2.1 (PI-RADS) category 3 lesions are a challenge in the clinical workflow. A better detection of the infrequently occurring clinically significant prostate cancer (csPCa) in PI-RADS 3 lesions is an important objective. The purpose of this study was to evaluate if feature maps calculated from T2-weighted (T2w) 3 Tesla (3T) MRI can help detect csPCa in PI-RADS category 3 lesions. In-house biparametric 3T prostate MRI examinations acquired between January 2019 and June 2023 because of elevated prostate-specific antigen (PSA) levels were retrospectively screened. Inclusion criteria were a PI-RADS 3 lesion and available results of an ultrasound-guided targeted and systematic biopsy. Exclusion criteria were a simultaneous PI-RADS category 4 or 5 lesion and hip replacement. Target lesions with the International Society of Urological Pathology (ISUP) grade group 1 were rated clinically insignificant PCa (ciPCa) and ≥2 csPCa. This resulted in 52 patients being included in the final analysis, of whom 11 (21.1%), 8 (15.4%), and 33 (63.5%) patients had csPCa, ciPCa, and no PCa, respectively, with the latter two groups being combined as non-csPCa. Eight of the csPCas were located in the peripheral zone (PZ) and three in the transition zone (TZ). In the non-csPCa group, 29 were located in the PZ and 12 in the TZ. Target lesions were marked with volumes of interest (VOIs) on axial T2w images. Axial T2w images were then converted to 93 feature maps. VOIs were copied into the maps, and feature quantity was retrieved directly. Features were tested for significant differences with the Mann-Whitney U-test. Univariate models for single feature performance and bivariate models implementing PSA density (PSAD) were calculated. Ten map-derived features differed significantly between the csPCa and non-csPCa groups (AUCs: 0.70-0.84). The diagnostic performance for TZ lesions (AUC: 0.83-1.00) was superior to PZ lesions (AUC: 0.74-0.85). In the bivariate models, performance in the PZ improved with AUCs >0.90 throughout. Parametric feature maps alone and as bivariate models with PSAD can (?) noninvasively identify csPCa in PI-RADS 3 lesions and could serve as a quantitative tool reducing ambiguity in PI-RADS 3 lesions.

Keywords: magnetic resonance imaging; prostate; prostatic neoplasms; prostatitis; radiomics.

Figure 1

Flow diagram of patients. PCa: Prostate cancer.

Figure 2

Workflow from the original images to the parametric feature maps. The original T2w images are computed into parametric maps using a script-based pipeline. A separate map is calculated for each of the 93 features, resulting in 93 feature maps per patient. As a parallel step, the target lesions are segmented manually in the original T2w images with volumes of interest (VOIs) using the open-source software 3D sclicer (www.slicer.org). In the software application, the axial (red window), coronal (green window), and sagittal (yellow window) reformation including the VOI is shown. In the next automated step, the script is used to copy the VOI into all the maps at the exact location as in the original image. The quantity of the respective feature can be determined by extracting the mean from the VOI.

Figure 3

Examples of parametric feature maps. On the left, the original axial T2-weighted (T2w) image and the segmentation, i.e., the volume of interest (VOI, yellow) of the target lesion are shown, and on the right, exemplary parametric feature maps are shown. The feature maps carry the same spatial information as the original image and reflect the feature quantity in brightness. Quantitative information on the feature can be directly drawn from the map once it is computed by extracting the mean. The VOI is copied into the feature maps in an automated step of the software script.

Figure 4

The steps in the data analysis are shown. Differences between the patients with clinically significant prostate cancer (csPCa) and patients without PCa and with clinically insignificant PCa, grouped as non-csPCa, were investigated. In the univariate analysis, significant differences between the two groups were analyzed using the Mann–Whitney U (MWU) test. Generalized linear models (GLMs) of the single map-derived features were built, and the performance was assessed using a receiver operating characteristic (ROC) curve analysis. In the bivariate analysis, data were upsampled, and GLMs combining the single map-derived features with serum prostate-specific antigen density (PSAD) were built applying ridge regression. Performance was assessed with ROC via bootstrapping.

Figure 5

Exemplary feature maps of best-performing features in the peripheral zone. In the upper row, images of a patient with clinically insignificant prostate cancer (ciPCa, Gleason Score 3 + 3, prostate-specific antigen (PSA) density: 0.17 (ng/mL)/cm³) and in the lower row images of a patient with clinically significant prostate cancer (csPCa, Gleason Score 3 + 4, PSA density: 0.17 (ng/mL)/cm³) are shown. On the left, the original T2-weighted axial image and the segmented target lesion marked with a volume of interest (VOI, yellow) are displayed. On the right, the parametric feature maps of the first-order median, first-order total energy, and glcm inverse difference normalized are shown. The map-derived features first-order median and first-order total energy revealed excellent diagnostic performance in differentiating csPCa from non-csPCa (no PCa and clinically insignificant PCa). VOIs were copied based on software in the maps, as shown for the glcm inverse difference normalized map, to extract feature quantity directly.

Figure 6

Feature maps with the best diagnostic performance in the transition zone. Both rows display images of a PI-RADS 3 lesion in the left anterior transition zone. Next to the original T2-weighted (T2w) images, the segmentation of the target lesion, i.e., the volume of interest (VOI, yellow) is shown. The patient in the upper row had a prostate-specific antigen (PSA) density of 0.09 (ng/mL)/cm³, and the biopsy revealed no prostate cancer (PCa). The lower row shows images of a patient with a PSA density of 0.16 (ng/mL)/cm³ and biopsy-proven clinically significant prostate cancer (csPCa) with a Gleason Score of 4 + 5. The features derived from the first-order 10th percentile and minimum map differed significantly between the csPCa and non-csPCa groups with excellent to outstanding diagnostic performance.

Figure 7

The heatmap of the single-feature performance separated by zone. Map-derived features that were significantly different between the clinically significant prostate cancer (csPCa) and non-csPCa (clinically insignificant prostate cancer and no prostate cancer) groups are listed on the left. Ten features differentiated the two groups across all prostatic zones with areas under the curve (AUCs) of 0.70–0.84. Considering only lesions localized in the peripheral zone, the ten exact features differed significantly but with slightly better diagnostic performance (AUCs: 0.74–0.85). When the analysis was limited to lesions in the transition zone, seven features differed significantly with excellent to outstanding diagnostic performance (AUCs: 0.83–0.10).

Figure 8

Areas under the curve (AUCs) of the generalized linear models combining the performance of feature maps and prostate-specific antigen density (PSAD) are summarized in a heatmap, considering PI-RADS 3 category 3 lesions in the peripheral and transition zones together. PSAD was implemented in the models as a continuous variable (AUC PSAD and feature maps) and with a cutoff of 0.15 (ng/mL)/cm³ and 0.20 (ng/mL)/cm³.

Figure 9

The heatmap shows the areas under the curve (AUCs) of the generalized linear models combining the map-derived features with the prostate-specific antigen density (PSAD), considering PI-RADS category 3 lesions only in the peripheral zone. Models were calculated with PSAD as a continuous variable (AUC PSAD and feature maps) and with a cutoff of 0.15 (ng/mL)/cm³ and 0.20 (ng/mL)/cm³.