Variability in the quality and interpretation of multiparametric MRI (mpMRI) prostate scans is a problem that confronts radiologists. These exams can be difficult to interpret, particularly for those who aren't expert genitourinary radiologists.
Because mpMRI is widely used, radiology software developers have sought to improve and eliminate the drawbacks, especially to reduce the risk of errors for less skilled MRI readers and nonradiologists. Radiology resident programs may benefit most from AI's early advantages. If a large portion of students entering the field of radiology already had a comfortable relationship two years ago with how leveraging technology can augment patient care and quality, according to the American College of Radiology, the prospect of residents using AI has certainly gained momentum in 2024.
Whether AI is used as an educational tool for radiology residents or an AI model is applied to prostate MR imaging quality, cancer risk mapping, or as an aid to specialists outside of radiology, this article highlights four examples of how AI development has multiplied the options for prostate imaging, prostate tumor identification and quantification, and prostate cancer management.
Variability in the quality and interpretation of multiparametric MRI (mpMRI) prostate scans is a problem that confronts radiologists. These exams can be difficult to interpret, particularly for those who aren't expert genitourinary radiologists.
Because mpMRI is widely used, radiology software developers have sought to improve and eliminate the drawbacks, especially to reduce the risk of errors for less skilled MRI readers and nonradiologists. Radiology resident programs may benefit most from AI's early advantages. If a large portion of students entering the field of radiology already had a comfortable relationship two years ago with how leveraging technology can augment patient care and quality, according to the American College of Radiology, the prospect of residents using AI has certainly gained momentum in 2024.
Whether AI is used as an educational tool for radiology residents or an AI model is applied to prostate MR imaging quality, cancer risk mapping, or as an aid to specialists outside of radiology, this article highlights four examples of how AI development has multiplied the options for prostate imaging, prostate tumor identification and quantification, and prostate cancer management.
1. Prostate MR quality
At Weill Cornell Medical College in New York City, Heejong Kim, PhD; Daniel Margolis, MD; and colleagues have pointed out that the problem of high variability in MRI image quality hinders comparative research and multicenter clinical trials, which are especially important given the popularity of mpMRI. Despite PI-RADS core criteria, substantial quality gaps persist, wrote Kim and colleagues in a European Journal of Radiology August 2023 article.
"If you don't have good quality to begin with, you can't have a quality product in terms of the report," explained Margolis in an interview with AuntMinnie.com. Margolis is professor of radiology at Weill Cornell Medicine and director of Prostate MRI at Weill Cornell Medical College. "We were looking at whether or not AI has a role for quality because we know that quality is an issue for prostate MRI."
The team of radiologists, radiation oncologists, and radiological imaging scientists at Weill Cornell Medical College examined opportunities and limitations of employing AI in prostate MRI quality standardization. Their studies included software for accelerated MR image acquisition and automated prostate segmentation:
- Accelerated MR image acquisition. When speeding up acquisition reduces scan time, benefits could come from reduced costs and fewer motion artifacts. Compressed sensing and parallel imaging (e.g., [image-based] SENSE and [k-space-based] GRAPPA) are two deep-learning (DL) techniques that are thought to be mature and that have been applied to clinical settings. These techniques take a sparsely sampled image and reconstruct the corresponding fully sampled image, the Weill Cornell team wrote. Though there may be limitations, studies have suggested no significant difference between conventional bpMRI and DL-based bpMRI which may be good news.
- Automated prostate segmentation. The emergence of the U-Net architecture has an advantage as a deep-learning technique, as it preserves the original image structure and the context, Kim and colleagues continued in their reviews, which were published in August 2023. Studies that have used the U-Net architecture for prostate gland segmentation, for example, have achieved Dice similarity scores of around 0.9. High-quality segmentation is significant for its potential for guiding targeted biopsies, planning surgery or other treatments, and measuring volumes for active surveillance of prostate cancer.
"The application of AI to prostate imaging is still relatively new, and the expectation is the improved performance will allow more men to defer biopsy," Margolis said.
2. Radiology resident education
Expert radiologists and first-year radiology residents are well aware of the difficulties with mpMRI, according to Italian researcher Ali Forookhi, MD, of the University of Rome. Forookhi and a team of collaborators studied the benefits of a semi-automatic commercially available AI-assisted software on interreader agreement in Prostate Imaging-Reporting and Data System (PI-RADS) scoring at different Prostate Imaging-Quality system (PI-QUAL) ratings and grades of reader confidence in mpMRI. The software used in the study was Quantib Prostate, now known as Saige Prostate (DeepHealth).
For this study, researchers found that in cases of poor image quality and low reader confidence, the prostate AI-assisted software was useful for novice first-year radiology residents who were assigned to urogenital radiology rotations on a regular basis.
In the related scientific publication in the European Journal of Radiology last year at the time of the European Congress of Radiology (ECR), the authors noted that radiology residency training programs were expected to incorporate AI-integrated clinical workflows to serve as a source of feedback and guidance for prostate lesion localization in contrast-enhanced prostate MRI. This study also pointed to the AI-assisted prostate MRI software for its impact on improving mpMRI reporting times among novice readers who used it.
3. Prostate cancer risk maps
Other functions of AI are designed to quantify and automate prostate lesion identification, including volume and severity, to improve prostate cancer diagnostic methods. Ultimately, one of the goals is to develop mpMRI prostate cancer-specific biomarkers.
The University of California, San Francisco (UCSF) produces 3,000-plus prostate MRIs per year and is among the academic centers where researchers are exploring AI for prostate cancer.
Quantitative mpMRI can improve targeting of biopsy and treatment as well as decisions on active surveillance versus treatment, according to Matthew Gibbons, PhD, from the department of radiology and biomedical imaging at UCSF, who with colleagues in the UCSF Department of Urology and the Department of Pathology generated prostate cancer risk maps they detailed last year in the journal Magnetic Resonance Imaging.
Gibbons, along with principal investigator Susan Noworolski, PhD, also from the department of radiology and biomedical imaging; Jeffry Simko, MD, PhD, from urology and pathology; and Peter Carroll, MD, in urology, assessed the prostate cancer maps they generated for the ability to detect and locate lesions 0.1 cc and larger, as well as quantify lesion volume and Gleason grade group 3 and higher versus lower as validated by histopathology.
Overall, quantitative results could "assist radiologists in their assessments by estimating lesion volume and grade," Gibbons and colleagues explained in their paper, adding that in addition to tumor load, prostate cancer masks can be used to extract mean MRI properties within cancer volumes. While the paper recognized the limitations of the early-development prostate cancer maps, quantities resulting from their method could be applied for population analysis of progression.
In an interview with AuntMinnie.com, UCSF's Thomas Hope, MD, said some expert radiologists may forego adding AI into the workflow when experienced readers are familiar with what to expect with prostate MR; however, lower volume practices may benefit from AI assistance. Hope is a professor of radiology and vice chair of clinical operations and strategy for UCSF Department of Radiology, where he also specializes in molecular therapy.
"There's only one role of prostate MR AI, which is to give you a map of is there a clinically significant cancer or not," Hope continued. "We have a massive shortage of radiologists. The biggest issue in radiology is a lack of FTE [full-time equivalent radiologists]. We don't have the people to do the work, and the work keeps growing at 10% per year, but the number of radiologists stays the same. What we need from AI is to make us more efficient ... read a study faster ... read 20 MRIs in the time it used to take to read five and handle the volume that's in front of me."
4. Usefulness outside radiology
While some radiologists may not be planning to add prostate AI tools into the clinical workflow soon, Hope said, other specialists may find that AI for prostate cancer practice management is particularly useful. This may be the case for urologists performing fusion biopsies where their focus is not metastases or nodes, he said.
Hope also predicted that radiation oncologists will find prostate AI tools useful, building off the results of the FLAME trial. As many AuntMinnie.com readers may know, the FLAME trial was a phase III, multicenter randomized control trial that tested delivering a focal radiotherapy boost to tumors visible on MRI. This intraprostatic focal boost is not routinely used, Hope said.
"What the FLAME trial did was show that you can do whole gland radiation therapy, or you can do that and then use the prostate MRI to boost where the tumor is and put extra radiation where the tumor is on the MRI, and they showed that when you use MR guided radiation therapy, boosting the region where the tumor is, you get better outcomes," Hope explained. "For prostate therapy planning, you could imagine prostate MRI aiding a radiation oncologist telling them where to boost."
Bottom line
The prevalence, complexity, and often long-length patient care and management of prostate cancer cases combined with workload pressure on expert radiologists make prostate MRI an easy target for AI-assisted software development.
AI tools for and beyond radiologists are expected to improve facets of prostate MR imaging across a range of tasks. Improving diagnostic accuracy for the detection of cancer and moving to noninvasive ways to predict prognosis and guide management for prostate cancer would improve patient care.
"The next thing that's going to impact prostate MR most likely is quantitative imaging," Margolis added. "The way we think about prostate MR is mostly qualitative ... and being able to measure the functional parameters and use that as a cutoff is going to be the next step."
Expert radiologists, urologists, and radiation oncologists will want to continue monitoring AI for its value as a quality assistant, for efficiencies gained, and for improving patient outcomes, in addition to bridging the experience gap for medical residents.