Prostate Cancer Differentiation and Aggressiveness: Assessment With a Radiomic-Based Model vs. PI-RADS v2

Abstract

Background: Multiparametric MRI (mp-MRI) combined with machine-aided approaches have shown high accuracy and sensitivity in prostate cancer (PCa) diagnosis. However, radiomics-based analysis has not been thoroughly compared with Prostate Imaging and Reporting and Data System version 2 (PI-RADS v2) scores.

Purpose: To develop and validate a radiomics-based model for differentiating PCa and assessing its aggressiveness compared with PI-RADS v2 scores.

Study type: Retrospective.

Population: In all, 182 patients with biopsy-proven PCa and 199 patients with a biopsy-proven absence of cancer were enrolled in our study.

Field strength/sequence: Conventional and diffusion-weighted MR images (b values = 0, 1000 sec/mm² ) were acquired on a 3.0T MR scanner.

Assessment: A total of 396 features and 385 features were extracted from apparent diffusion coefficient (ADC) images and T₂ WI, respectively. A predictive model was constructed for differentiating PCa from non-PCa and high-grade from low-grade PCa. The diagnostic performance of each radiomics-based model was compared with that of the PI-RADS v2 scores.

Statistical tests: A radiomics-based predictive model was constructed by logistic regression analysis. 70% of the patients were assigned to the training group, and the remaining were assigned to the validation group. The diagnostic efficacy was analyzed with receiver operating characteristic (ROC) in both the training and validation groups.

Results: For PCa versus non-PCa, the validation model had an area under the ROC curve (AUC) of 0.985, 0.982, and 0.999 with T₂ WI, ADC, and T₂ WI&ADC features, respectively. For low-grade versus high-grade PCa, the validation model had an AUC of 0.865, 0.888, and 0.93 with T₂ WI, ADC, and T₂ WI&ADC features, respectively. PI-RADS v2 had an AUC of 0.867 in differentiating PCa from non-PCa and an AUC of 0.763 in differentiating high-grade from low-grade PCa.

Data conclusion: Both the T₂ WI- and ADC-based radiomics models showed high diagnostic efficacy and outperformed the PI-RADS v2 scores in distinguishing cancerous vs. noncancerous prostate tissue and high-grade vs. low-grade PCa.

Level of evidence: 3 Technical Efficacy: Stage 2 J. Magn. Reson. Imaging 2019;49:875-884.

Figure 1

Flow diagram of patient selection.

Figure 2

ROI delineation of noncancerous tissue and low‐ and high‐grade PCa. (a,d,g: T₂WI images; b,e,h: ADC images; c,f,i: pathological images.)

Figure 3

The importance of features extracted from T₂WI, ADC, and T₂WI&ADC images to distinguish PCa from noncancerous patients is shown in A–C, respectively.

Figure 4

The importance of features extracted from T₂WI, ADC, and T₂WI&ADC images to distinguish high‐ from low‐grade PCa is shown in A–C, respectively.

Figure 5

ROC curves for radiomics‐based ADC, T₂WI, and ADC&T₂WI model and PI‐RADS score performance in distinguishing PCa vs. non‐PCa in the training and validation groups, respectively.

Figure 6

ROC curves for radiomics‐based ADC, T₂WI, and ADC&T₂WI model and PI‐RADS score performance in distinguishing GS 6 vs. GS ≥7 PCa in the training and validation groups, respectively.