Magnetic Resonance Imaging (MRI) for Prostate Cancer Diagnosis

Multiparametric MRI is being increasingly used for prostate cancer (PCa) screening and identifying targets for biopsy. Several large studies including the PROMIS trial have established multiparametric MRI to have a significantly high negative predictive value in identifying clinically significant (cs) PCa.¹ However, MRI is limited by moderate specificity and inter-reader variations in interpretation of PCa (Prostate Imaging–Reporting and Data System [PI-RADS™]). Consequently, biopsies continue to be the standard of care for diagnosis of csPCa.

Artificial Intelligence (AI) for Risk Stratification of PCa on MRI

AI-based approaches are being widely explored, addressing the limitations of MRI for improving noninvasive diagnosis of PCa. Radiomics, which involves computational extraction of features corresponding to subvisual patterns of underlying tissue heterogeneity on imaging, has shown promising results in characterizing PCa on MRI. In our recent study, we explored the tumor along with its surrounding habitat (peritumoral region of PCa) in terms of radiomic signatures from T2-weighted and diffusion-weighted MRI sequences (fig. 1).² We observed that our radiomics model resulted in >8% improvement in AUC in terms of risk stratification of PCa compared to PI-RADS v2.0 in a multi-institutional cohort (231 patients). These features were likely capturing the differential heterogeneity between low- and high-risk lesions both within and surrounding the tumor on MRI, which was also verified by their correlations with histomorphometric attributes on corresponding pathology.

Figure 1. Radiomic feature maps overlaid on peri-tumoral region of prostate cancer lesions on T2W MRI showing differential heterogeneity captured between low and high-risk lesions as defined by the D’Amico criteria.

Deep learning-based approaches that use sophisticated neural networks to train models without the need for extracting specific imaging biomarkers have also been explored in the context of prostate MRI. For instance, we developed a deep learning-based radiomic risk score for identifying csPCa on MRI in a recently published study.³ The predictions from this deep learning model were integrated with clinical parameters including prostate specific antigen, and prostate and lesion volume to build an integrated nomogram to identify csPCa in a large multi-institutional cohort (592 patients). We showed that deep learning derived image patterns from the tumor and the surrounding peritumoral regions were able to characterize csPCa and outperformed PI-RADS v2.0-based assessment of csPCa by 4% in terms of AUC. This model was also prognostic of biochemical recurrence in a subset of patients who underwent radical prostatectomy with a median followup of 3 years. These results and other studies demonstrate the value of AI-based approaches in capturing subtle, subvisual patterns on MRI and potentially improve risk stratification of PCa.

Promise of AI for Population-Specific Models

However, there is a significant disparity in PCa incidence and mortality between different population groups. Specifically, African American (AA) men have a 1.5 times greater likelihood of developing PCa than Caucasian American (CA) men and are 2.2 times more likely to die of it.⁴ Beyond socioeconomic factors, biological and phenotypic differences are observed to exist between men of different populations. AA men with PCa tended to show significant overexpression of genomic and molecular markers associated with DNA damage, hypoxia, apoptosis, inflammation and immune response compared to CA men in a large (1,152 patients) multisite study.⁵ These differences observed at the genomic and molecular scale will persist at the morphological scale that can be discernible on radiology⁶ and pathology⁷ images. Most of the clinical nomograms and diagnostic assays so far do not incorporate population-specific differences and were designed largely based on CA men.

Given the recent promise of AI in improved characterization of PCa on imaging, there is a need and an opportunity to explore population-specific AI approaches for PCa risk stratification on MRI. In another recent study, we observed that computationally derived image features (pathomics) of cancer nuclei from stromal regions of PCa on digitized prostatectomy specimen are prognostic of recurrence in AA men and outperformed a population agnostic model.⁸ In a preliminary cohort of 41 PCa patients with biopsy-confirmed ISUP (International Society of Urological Pathology) Grade Group 2 PCa, radiomic features of PCa lesions on apparent diffusion coefficient maps showed statistically significant differences (p <0.05) between AA and CA patients when none of the other clinical variables showed significant differences between the population groups (fig. 2). These preliminary results presented at AUA2020 suggest the potential of AI in identifying population-specific heterogeneity on MRI.⁹

Figure 2. Radiomic feature maps (Haralick and CoLlAGe) overlaid on ADC maps of prostate cancer ROIs (left) along with corresponding distribution plots (right). AA patients tend to have a relative overexpression of radiomic features compared to CA men.

Looking Ahead

AI-based population-specific risk stratification approaches using MRI can potentially allow for early identification of PCa in high-risk populations, specifically AA men. This may also enable development of effective, imaging-based noninvasive approaches for monitoring PCa progression in AA men within the active surveillance setting and preemptively identify those with adverse outcomes post-radical therapy. One of the primary challenges in developing and bringing population-specific approaches to clinical deployment is to isolate factors,¹⁰ including socioeconomic determinants, access to health care, other comorbidities and therapies. These biases need to be addressed up front while also tackling issues of generalizability, reproducibility of AI models, and robustness to site and scanner-specific variations in imaging. Significant efforts are needed in curating large multi-institutional cohorts of men with PCa from diverse population groups. Enforcing strict controls for social determinants of health, AI could pave the way in quantifying the biological and genetic heterogeneity between different population groups on imaging and allow for building imaging-based, population-specific PCa risk stratification approaches.