Read the Full Video Transcript

Andrea Miyahira: Hi, everyone. Thank you for joining us today. I'm Andrea Miyahira at the Prostate Cancer Foundation. With me today is Dr. Tamara Jamaspishvili, an assistant professor at SUNY Upstate Medical University. She'll be discussing her group's recent paper, "Artificial Intelligence-Based PTEN Loss Assessment as an Early Predictor of Prostate Cancer Metastasis After Surgery: A Multicenter Retrospective Study." This was published in Modern Pathology.

Thanks so much for joining me today, Dr. Jamaspishvili.

Tamara Jamaspishvili: Thank you so much, Andrea, and thank you for the invitation. It's really my pleasure to present the results of our study today and share with the audience. I'm going to talk about our recent publication in Modern Pathology about our recent work on PTEN loss assessment in prostate cancer. The name of this publication is "Artificial Intelligence-Based PTEN Loss Assessment as an Early Predictor of Prostate Cancer Metastasis After Surgery: A Multicenter Retrospective Study."

I want to give you a little bit of an overview on PTEN loss. I'm pretty sure, to this audience, probably this biomarker is well-known. The PTEN, as we know, is a tumor suppressor, so it's a recognized marker of aggressive prostate cancer. It is located on chromosome 10 and, when it is lost, which mostly occurs due to genomic deletion, it over activates the AKT pathway, which stimulates dysregulation of the growth and migration and invasion of the cells.


It's known that, in primary prostate cancer, the loss of PTEN happens in around 20% of the cases, and in metastatic settings, it's around 40 to 50%, so it increases. Many studies, including ours, have shown that PTEN loss is associated with adverse pathology and bad clinical outcomes, including upgrading at radical prostatectomy, developing biochemical recurrence, also castration-resistant prostate cancer, and metastasis.


These are some publications that we published on this topic. Historically, PTEN was evaluated using fluorescence in situ hybridization, FISH, which was considered as the gold standard for assessing genomic loss, but because it's very labor-intensive and it's very difficult to incorporate and implement in clinical, especially in the research setting, immunohistochemistry (IHC) fairly recently became an alternative. Dr. Tamara Lotan at Johns Hopkins University developed a genetically-validated test, and then we collaborated with our group and participated in the optimization and validation of the immunohistochemistry assay.


PTEN testing is now used as, I mentioned, an alternative to FISH testing. The only caveat with IHC is that the loss of this gene is very heterogeneous and is non-uniform basically at the protein level, which makes it very difficult to interpret and correlate the PTEN loss with the prognosis and outcome. For example, in this figure, as you can see, the survival curve, we're showing the cases, the patients, who have partial PTEN loss and homogeneous PTEN or complete PTEN loss and the ones who do not have PTEN loss. The ones who have complete PTEN loss in all cancer cells in their tissues are clearly experiencing the worst outcome compared to the cases with the partial loss.


It was really very clear that, unfortunately, all the methodology and assessment of PTEN was manual. As in most laboratories for most of the biomarkers still, the assessment is basically done by the pathologist using the microscope, so it was very subjective, and we thought that a proportion of these cases who were experiencing heterogeneous PTEN loss would need to be better classified to make this precise correlation with the outcome. Then we hypothesized basically that quantitative measurements of PTEN loss would probably provide a better prognostic relevance compared to the manual assessment.


To test this hypothesis, my group and I decided to design a proof of concept study utilizing commercial digital image analysis software to annotate the scanned immunohistochemical images and quantify the number of the cancer cells with the loss. It wasn't an AI-based study, so we used just a third-party commercial software just to simply quantify the number of the cells and to show that there was a certain clinical amount or clinically-relevant amount of the cancer cells that would be better associated with the outcome.


Indeed, as you can see from the Kaplan-Meier curves, using the quantitative digital pathology-based approach, we were able to identify more cases with early biochemical recurrence than we would do by using manual scoring. That really encouraged us to move to the next project, which was the continuation, and we teamed with the National Cancer Institute, Dr. Stephanie Harmon, who led basically AI and computer science part of the entire project, and she helped us to design and create the AI-based, fully automated in-house algorithm for PTEN detection.


We published the second paper also in Modern Pathology. It was basically the methodology paper. We were successfully able to train and validate on the independent tissue microarray cohort. This algorithm that would predict the cancer and as well the TMA, tissue microarray cores with the cancer as well as the PTEN loss into those cores as you can see from the heat maps.

The other, the third project which we also continued working on was the same group. It was a validation study which we did on the Canary Foundation. It was a seven multicenter cohort and, again, with Dr. Stephanie Harmon's group. For the first time, we were able to validate and show the clinical value in relation to the metastasis prediction. I can show these results on the following slides.

This image here shows the overall workflow development. As you can see, the first part, the workflow runs on the tissue microarrays and basically it computes and gives the probability of the cancer and probability of the PTEN loss for each core. The second part of the entire workflow, which in turn also contains the multiple steps, it's a patient-based evaluation where, here, we basically algorithm analyzes individual patients and computes the clinically actionable and clinically relevant threshold. We call the high-risk qPTEN as a name for this clinical threshold.


This is the more in-depth view of the entire workflow. As you can see, it consists of almost seven steps, and so the part of this work flows from step two to five basically. It's an AI-based algorithm prediction where, again, the algorithm computes the clinically relevant threshold, and the next or the last steps, six and seven, is a statistical modeling where the algorithm runs on the entire specimen from the individual patients to derive this threshold. In the table, we're just showing how we successfully validated across the seven centers using a logistic regression analysis, so these clinically-relevant PTEN amounts.


That was the composition of the Canary Foundation retrospective TMA cohort. The good thing with the cohort was that, first of all, it was a pretty large cohort of 1,025 patients. These patients were distributed across retrospective TMAs, and we stained all these TMAs at Queen's University while I was doing this study or started this study.


The good thing was that the entire cohort also contained or had the patient with the metastasis. It was really very interesting for the very first time to see whether, clinically, our developed algorithm had any relation to the metastasis prediction. As you can see from this Kaplan-Meier curve, indeed, the high-risk AI-qPTEN demonstrated a statistically significant association with the shorter metastasis for survival compared to the manual scoring. Similarly, in multivariable analysis adjusted for CAPRA, surgical CAPRA, we showed also superior prognostic performance of higher risk AI-qPTEN in relation to the metastasis for survival and the recurrence-free survival compared to the conventional PTEN.


Another table here, we show that interestingly we analyzed this subset of the patients who had no adverse post-surgical clinical features, in other words, the patients who would classify according to the CAPRA as a low risk because they didn't have any lymph node invasion, extra prostatic invasion, seminal vesicle invasion or positive surgical margin. We specifically picked these patients, 295 cases, to look whether the AI-qPTEN could identify anything, and then, indeed, as you can see, we found that our thresholds also maintain statistical significance compared to the manual assessment of PTEN.


In this decision curve, lastly, we wanted to basically assess the net benefit increase of including high risk in clinical risk stratification tools. As you can see, this purple curve here is a combined model that involves high risk AI-qPTEN with CAPRA-S, and it showed the greater net benefit of treating this patient postoperatively which were diagnosed with prostate cancer compared to using the CAPRA-S alone.


Overall, this multicenter study demonstrates the potential of improving post-surgical risk stratification of patients with prostate cancer using quantitative AI-based approach as a use basically of single molecular digital biomarker test in combination with existing risk stratification tools such as CAPRA.

Also, we believe that, as a biomarker-guided trial becoming more common, it will be very interesting to assess the clinical value of this workflow in some prospective clinical trials to better understand the benefit, for example, postsurgical intensified treatments in a prostate cancer patient and, obviously, assess the predictive value of this algorithm as well.


This project was honored to get the award from Digital Pathology Association last year. This project basically was the best research award at the conference, so we're very proud of it and most important, of course, the project based on all this preliminary studies was funded by the Prostate Cancer Foundation.


I was selected as a young investigator award to class 2022, and now we are trying basically to assess the clinical value and to continue work on this topic and assess the clinical value of AI-qPTEN in a predictive context, so, specifically, the questions we will be asking is that whether this workflow can predict, for example, duration of androgen deprivation therapy in high and very high-risk patients as well as predict, for example, development of castrate-resistant prostate cancer.


In addition to qPTEN, we also are trying to incorporate the other markers which we know may have a predictive value such as the p53 and Rb1 in the model. This is our current ongoing project on this topic.

Thank you very much again for your attention.


Andrea Miyahira: Thank you so much for sharing this presentation with us. Just a few questions, is the AI algorithm only looking at PTEN staining or does it also incorporate any morphological features?

Tamara Jamaspishvili: The interesting and very unique of the... and that's why we got even more proud of our results and discovery that this only assess and looks at the immunohistochemistry, so the basically protein expression of solely PTEN. Yeah, our algorithm incorporates as well the cancer detection part, but this is only just the cancer that we did not train the algorithm purposely on the H&Es to predict, for example, or find the characteristic unknown morphological features. This could be a next step, and we are working with our collaborators who are interested to incorporate, for example, AI-qPTEN in relation to the morphological assessment of the prostate that could have even a better clinical decision power for the prostate cancer patients.

Andrea Miyahira: How variable is the PTEN IHC staining across institutions? Can the AI algorithm compensate for these technical differences?

Tamara Jamaspishvili: We used the stain normalization process as a part of the algorithm. The stain normalization is very important step in image analysis and image processing. There are different methods obviously of stain normalization. In our particular study, we used the Macenko method. The idea of the method is that it averages basically all stain differences coming from different analytical variables, for example, using a different type of antibodies and creates the referenced average matrices and, what we did, we basically used independent cohorts from single institution as a reference and we normalized multi-center staining right from the multi-center cohort against this reference.

Interestingly, when we compared also performance of the algorithm and especially on the cases which were unequivocal and very difficult for pathologists to evaluate, we saw a clear difference, so the inaccuracy. For example, the accuracy pre-normalized images was very difficult by pathologists to assess and post-normalization accuracy was definitely higher. Yes, we incorporated certain methods to overcome that issue.

Andrea Miyahira: Wonderful. How do you envision this tool being integrated into clinical workflows at non-academic institutions?

Tamara Jamaspishvili: Yeah. In a clinical workflow, I would imagine, first of all, because again this is a digital pathology tool, we have to consider many steps and the various infrastructural requirements, so, therefore, digital pathology such as you simply need... The facility has to have a scanner at the very simplest way, but logistically I would or clinically as well, so from clinical standpoint, I would envision using this workflow that, let's say, the patient comes for prostate cancer to the urologist and urologist and the surgeon does the surgery and sends specimens to the pathology lab and a pathologist basically chooses the best representative slide, the one slide from the surgical specimen, and runs this algorithm and confuses the high-risk score. This high-risk score of AI-qPTEN is reported basically back to the urologist, and the urologists make its own decision whether to incorporate or use any further treatment for this patient to avoid further metastasis in the future.

Andrea Miyahira: Thank you so much for sharing this wonderful study with us today. I look forward to your next study results.

Tamara Jamaspishvili: Thank you so much. I appreciate it. Thank you for inviting me.

Andrea Miyahira: Thank you.