PCa Commentary | Volume 183 – November 2023

Posted by Edward Weber | November 2023

Artificial Intelligence – The Wave of the Future in Prostate Cancer Diagnosis, Risk Stratification and Prediction of Outcomes

Background

A forceful sea change is roiling in the management of this disease: artificial intelligence (AI) is coming into prominence benefiting cancer diagnosis, risk stratification, and the prediction of response to therapy for individual patients. Multimodal AI (using deep learning) draws upon increasingly large stores of clinical and outcome data and digital histopathology to identify patterns that can predict therapeutic benefits for personalized treatment. The reference datasets might be extensive information generated from large clinical trials, a repository of annotated pathology specimens or imaging datasets (such as those of MRIs). In recent years, the literature has reported an impressively extensive body of studies of AI applications. This Commentary will offer four examples that could inform clinical practice.

Guiding the Use of Androgen Suppression Adjuvant to Radiation Therapy:

Two published pace-setting studies on this subject were covered in previous PCa Commentaries.

PCa Commentary #180: Dr. Dan Spratt and colleagues (NEJM Evidence, June 2023) developed an AI program identifying a biomarker, derived from 1719 patients in 5 large clinical trials, that predicts the marginal benefit of adding short-term hormone therapy to primary radiation for a man with localized prostate cancer. By applying the AI-derived model, the 15-year estimate for developing distant metastases for biomarker-positive patients treated with ADT was 4% v 14.4% for biomarker-negative ADT patients. This test is commercially available at ArteraAI Prostate Test where it is fully explained.

PCa Commentary #181: The second study was presented in abstract form by Armstrong et al. at ASCO 2023:
“Development and validation of an AI-derived digital pathology-based biomarker to predict benefit of …[28 months vs 4 months] androgen deprivation therapy with radiotherapy in men with localized high-risk prostate cancer … .”

Improving Cancer Detection Based on Prostate MRI

Multiparametric MRI functions to detect prostate cancer and guide MRI-ultrasound fusion biopsies and treatment. Optimally an MRI study leads to a biopsy of only clinically significant cancer, defined as Gleason score of 7 or greater. The PIRADS (Prostate Imaging-Reporting and Data System) offers a numerical estimate for the likelihood of the presence of cancer based on the MRI. This estimate of cancer’s aggressiveness is based on a scale of 1 through 5, with 1 and 2 representing low likelihood, 3 – concerning and 4 and 5 very suspicious for cancer. Unfortunately, even this system has deficiencies.

As reported by Bhattacharya et al, Ther Adv Urol. 2022, compared to the examination of a  subsequent prostatectomy as ground truth, “12% of aggressive cancers, mostly those less than 1 cm, were missed on MRIs. False positive rate was greater than 35%. Additionally, there is high inter-reader variability. As a result, many unnecessary biopsies continue to be performed”.

“Radiomics” refers to the analytic system for extracting features of significance from MRI images that are not perceived by visual examination. There is extensive research on using radiomics to facilitate greater interpretive and predictive information from the MRI and to improve on the stratification of the PI-RAD system.

The underestimation of cancer extent on MRI is addressed by Priester et al, in ”Prediction and Mapping of Intra-prostatic Tumor Extent with Artificial Intelligence”, European Urology June 2023. One of their objectives was to improve the accuracy of tumor delineation of clinically significant cancer on the MRI to facilitate more effective focal radiation therapy, a technique increasingly gaining acceptance for intermediate-risk cancer. To accomplish this an AI model was developed combining MRI imaging, biopsy data, PSA and PSA Density values and prostate volume to produce “three-dimensional cancer estimation maps and margins.” When compared to cancer extent found on companion prostatectomy specimens, the AI estimate model was more accurate than one based on conventional MRI interpretation, ensuring a better outcome for focal radiation therapy.  This AI model is FDA-approved and commercially available: “Avenda Health AI Prostate Cancer Planning Software”.

Predicting Early Recurrence after Prostatectomy

An AI-powered method was developed to predict early recurrence at 36 months after prostatectomy based on digitized pathology slides in conjunction with clinical and outcome data (Huang et al, JCO Clin Cancer Inform, 2022). The AI platform was trained on 243 digitized whole-mount slides of prostatectomy specimens combined with information about Gleason score, staging, margin status and clinical outcome. The method was validated on 92 patients who had recurrence in <3 years and 151 who recurred after 3 years.

The 100,000 x 100,000  pixels surveyed per slide has the advantage (compared to the standard microscopic evaluation) of recording small, but relevant, regions that escape visual notice, and perhaps most important, capture immune features (not included in the Gleason Grade Groupings), of the tumor microenvironment and stroma (supporting cells) that are so influential in driving the cancer’s behavior.

Based on this data the study provided a prediction of biochemical recurrence within 3 years after surgery for men with cancer in all Gleason Grade Groups and performed better than conventional risk stratification systems. The authors felt that this AI method for identifying patients at risk for early recurrence would benefit the selection of personalized treatment.

AI in Association with Cancer Diagnosis and Gleason Grading

AI algorithms have performed well in this task and have been multiply validated. Many are certified for clinical use. The usefulness of AI has not been to replace pathologists in analyzing biopsy specimens, but to augment the heavy workload of pathologists. “Pathologic examination of prostate specimens is laborious and time-consuming due to the large number of slides per case – 50-100 slides per case.” (Tolkach)

In a comprehensive study by Tolkach et al. (Nature NPI Precision Oncology, 2023) 7473 biopsy cores were digitized, assessed for tumor detection and the results were compared with the findings of expert pathologists. “We show high levels of diagnostic accuracy for prostate cancer detection and agreement levels for Gleason grading comparable with experienced genitourinary pathologists.”

BOTTOM LINE

The application of artificial intelligence is becoming widespread in the field of prostate cancer with the promise of improving clinical practice.

“We appreciate the unfailing assistance of the librarians at Providence/Swedish.”