Prostate cancer is the second most common solid organ cancer among men worldwide. In the U.S., it is also one of the leading causes of cancer-related death in men.
Keshavarz
Traditionally, clinically significant prostate cancer (csPCa) has been detected by serum PSA density level (total PSA divided by prostate volume) and PSA velocity (PSA measurement increase over time) supplemented by a digital rectal exam and template transrectal biopsy. However, each of these tests have limitations that may lead to significant errors in diagnosis.
In fact, when combined, it is estimated that only 50% of patients have a truly accurate diagnosis.
“Better detection of csPCa would mean more accurately triaging patients to active surveillance and partial or whole gland therapy,” said Pedram Keshavarz, MD, a postdoctoral research fellow at UCLA.
This is why Dr. Keshavarz said radiologists have largely turned to multi-parametric magnetic resonance imaging (mpMRI) based approaches.
“Today, mpMRI is the most widely available and proven imaging modality for the detection, grading and staging of prostate cancer, especially for prostate biopsy planning,” he said.
While beneficial, mpMRI is far from perfect. According to one study (Bajgiran et al.), it can still miss up to 20% of csPCas. The study, which looked at the characteristics of missed PCa lesions by mpMRI on 518 patients with 1,085 PCa lesions, found that the missed lesions tended to be small and non-cohesive compared to the detected lesions.
According to Dr. Keshavarz, AI could help reduce mpMRI’s miss rate.
Helping Radiologists Accurately Diagnose Prostate Cancer
Speaking at a Wednesday session, Dr. Keshavarz noted how a variety of machine and deep learning AI techniques are now being used to detect and classify csPCa using qualitative and quantitative features of mpMRI imaging exams.
“The detection stage is so important for biopsy planning, especially for targeted biopsy, and results like these show that AI may be useful for estimating the probability of csPCa and planning,” Dr. Keshavarz noted.
To better understand the potential of AI in the field, Dr. Keshavarz led a study that systematically reviewed 26 published studies in 6,231 patients with PCa.
“Our aim was to investigate the overall performance of AI models and calculate a pooled performance of AI algorithms for detecting and classifying PCa on mpMRI,” he explained.
What they found was that, although AI-based algorithms need to train on large and diverse populations to minimize bias, AI-based models that use imaging features show good performance for detecting or classifying prostate cancer and can help radiologists accurately diagnose csPCa.
“For some specific applications, such as tumor detection in the transition zone or detection of csPCs among PIRADS 3 lesions, AI-based methods might even be superior to radiologists’ performance,” Dr. Keshavarz concluded. “Therefore, AI-based approaches could enable junior or less well-trained radiologists to better read and more accurately report prostate-MRI findings.”
Access the presentation, “Detection and Classification of Prostate Cancer on Multiparametric MR Imaging Using Artificial Intelligence: A systematic review of 6,231 Patients,” (W1-SSGU05-3) on demand at Meeting.RSNA.org.
