Improving Prostate Cancer Trial Imaging through AI: Part 2

Connect with Perceptive’s oncology imaging scientists at ASCO 2025 to discuss your clinical trial imaging needs.

Click Here to Schedule a Meeting

PSMA-PET Imaging: Looking Forward

In the previous installment to this series, we reviewed the groundbreaking VISION study that ushered in a new era of artificial intelligence and machine learning use in clinical trial imaging. During the study, Perceptive’s automated segmentation methodology used information from both PSMA-PET imaging as well as CT scans to mask the skeleton and reliably delineate bone lesions, ultimately helping the sponsor demonstrate that the whole-body SUV mean on PSMA-PET was predictive of survival.

Perceptive has since updated its automated segmentation methodology. Version 2 automatically delineates between individual lesions and automates all tissue types for Ga68 PSMA mCRPC disease. This enables:

• Automatic segmentation of both bone and soft-tissue lesions, significantly reducing human processing time
• Tunable post processing for flexibility across different disease stages or tracers
• Robust and dynamic thresholding, and flexibility to treat body areas differently, ex. rib lesions v. lumbar lesions
• Anatomical location labeling

Here we talk with Dr. Phil Kuo, first author for the two FDA-mandated sub-studies for VISION, to get his perspective on V2 of Perceptive’s methodology, the impact it will have on PSMA-PET imaging, and how he envisions AI-enabled image analysis will advance prostate cancer research and clinical care moving forward.

What do you see as the biggest advantages of Perceptive’s automated PSMA segmentation tool V2?

Dr. Kuo: Some critical modifications have been made in Version 2, the most important is that soft tissue disease can now be automatically segmented.

In Version 1, there were some areas we knew would be difficult to analyze, particularly where there’s curvature of the bones, like in the skull near the vertex and in the ribs – both of which are difficult for both humans and AI to segment.

It’s exciting that now, all disease can be segmented, even soft tissue disease.

Why is the auto segmentation of all disease important for prostate cancer research?

Dr. Kuo: This is important for multiple reasons:

1. Efficiency – In the end the most efficient processes win, and that’s what V2 delivers. By automating all disease segmentation, you limit the amount of manual analysis that’s typically required for soft tissue disease. (For reference, in the VISION trial, there were over 680 scans that had to be manually read for identification/analysis of soft tissue disease). Of course, you’ll always want an expert PSMA-PET reader to check these lesions, but with V2 their time can be significantly reduced to just the unusual circumstances and outliers that the AI algorithm hasn’t been trained on. This will amount to tremendous time savings, which I think most will agree is a significant advantage.
2. Reliability / Reproducibility – We all appreciate that in pharmaceutical development, reproducibility of results is critical. V2 of Perceptive’s methodology enables reproducibility not only across all trial participants, but also across the entire patient population. Which leads me to my next point.
3. Expansion – For the VISION trial, PSMA-PET scans were only performed for baseline / eligibility. Since then, additional research has added to our knowledge base about its value in clinical research. As we learn more, PSMA-PET will be used to track disease progression. In fact, PCWG4 is going to lay out a plan for using PSMA-PET in clinical trials, so I can see V2 of Perceptive’s segmentation algorithm being an important part of the process for analyzing serial follow-ups and determining patient response to therapy.

What do you see in the future for AI and ML in clinical practice?

Dr. Kuo: PSMA-PET imaging is a superior modality compared to conventional imaging, i.e., CT and bone scans. So, as we venture into using it for serial scans, we must do something with all the data it will gather, and quantitative analysis is the way to go.

Multiple trials have shown that tumor volume on serial follow-ups is critical to measuring response. It’s not SUVs, particularly with PSMA targeted agents.

As you can imagine, taking the model of PSMA targeted radioligand therapy, tumor cells and even adjacent cells that have the highest PSMA expression will get the highest dose of radiation most likely respond early. But there are scenarios where you kill off those tumor cells and you’re left with lower expression cells that may continue to grow.

So, it’s that change in volume that’s even more predictive of response than actual SUV means when you’re assessing serial PSMA-PETs. But no one does this in real world practice right now because it requires a predefined, reproducible methodology. And that’s where the updated segmentation tool will be critical. Of course, we first need to prove that it’s implementable, useful, and accurate in clinical trials before we roll it out into clinical practice.

But I think this is where we’re heading in the future.

Perceptive Imaging’s AI team is dedicated to advancing clinical trial imaging for optimal outcomes. Learn how they can help you get more out of your clinical trial imaging. Get in touch today.