Read the Full Video Transcript

Andrea Miyahira: Hi, everyone. I'm Andrea Miyahira at the Prostate Cancer Foundation. Here with me is Dr. Cornelia Ding, an assistant professor at UCSF. She will discuss her recent paper, "Predicting Prostate Cancer Grade Reclassification on Active Surveillance using a Deep Learning-based Grading Algorithm," that was published in JNCI. Dr. Ding, thanks so much for joining us today.

Cornelia Ding: Thank you for having me. I'm very excited to share this work from when I was a pathology fellow at Johns Hopkins University, working with Dr. De Marzo and Dr. Lotan, about predicting prostate cancer grade reclassification using AI.

This is collaborative work from many people with different specialties. So, of course, pathologists are the major drivers of this study, but also our urologists, oncologists, and our collaborator in India, AIRA MATRIX, providing the AI algorithm evaluating histology slides, are involved in this work.

Here is the disclosure. The study in this presentation is supported by a grant from the Prostate Cancer Foundation, sponsored by AIRA MATRIX. And, importantly, AIRA MATRIX did not assist with data analysis or writing of this study.

Clinical management for localized prostate cancer is strongly driven by pathology. By interpretation of prostate biopsies, we pathologists provide a diagnosis of prostate cancer as well as grade group, which is based on evaluation of tumor architecture using the Gleason grading system under the microscope.

The tumor architecture has been shown to be strongly associated with prostate cancer patient outcomes. As shown here, the Y-axis is biochemical recurrence-free survival, and the X-axis is time. You can see patients with grade group 1 disease have indolent disease progression over time. So they usually do not benefit from immediate definitive treatments such as radical prostatectomy. Many patients nowadays with grade group 1 disease stay on active surveillance.

However, a minority of these patients with low-grade disease may have disease progression over time. So our main question is, can we refine our pathology evaluation of prostate biopsies to better identify patients who may have disease progression? And can AI, artificial intelligence, help us here?

With collaboration from AIRA MATRIX, they developed a deep learning-based algorithm called AIRAProstate for tumor detection and Gleason grading using hematoxylin and eosin stained pathology whole-slide images.

In this particular study, we evaluated two different cohorts at Johns Hopkins University. The first cohort is a case-control design involving 138 patients with grade group 1 disease on diagnosis. Sixty-four patients later had grade reclassification, so they are the cases. Seventy-four patients stayed on active surveillance without any grade reclassification for more than eight years. So they are the controls.

One H&E slide per patient is regraded by both contemporary uropathologists and the AI algorithm AIRAProstate.

On the right side, you can see a head-to-head comparison from contemporary pathologist grading versus AIRAProstate. Pathologists and AIRAProstate largely agree upon the detection of cancer on each slide with 95% agreement. However, because these patients were enrolled many years ago and the grade system has been updated, the contemporary pathologists do not necessarily agree with the initial diagnosis of grade group 1.

Some of them are upgraded into grade group 2 or higher. However, a percentage are not significantly different. In contrast, upgrading by AIRAProstate is different between case and control, where 33% of cases are upgraded, while 8% of controls are upgraded. So you can see the odds ratio is 3.3 with P 0.04.

A different cohort is also evaluated. This is a little different because all the patients had MRI prior to the initial prostate biopsy, and the majority of the patients were enrolled after 2014. There are 169 patients with grade group 1 disease in this cohort; 94 of them stayed on active surveillance without grade reclassification during follow-up. However, 75 patients had later grade reclassification. Thirty-five of these patients underwent radical prostatectomy. H&E slides with cancers are evaluated by the AI algorithm AIRAProstate. You can see here substantial cases also get upgraded by the AI algorithm. Eighteen percent were upgraded by AI. However, 40% of the cases that had later grade reclassification were also upgraded. And these are different by hazard ratio 1.7, P equals 0.3.

In summary, this is the first study in the literature comparing a deep learning-based prostate cancer grading algorithm to clinical outcomes. We compared two different active surveillance cohorts at Johns Hopkins, both showing that the upgrading of initial biopsy by AI is associated with rapid or extreme grade reclassification, which predicts later disease progression.

So a take-home message from this study is that we have been using histopathology as biomarkers and also as the gold standard of diagnosis. This has provided important information that drives clinical decisions, especially in managing localized prostate cancer. Nowadays, with help from artificial intelligence in pathology, we could potentially refine this biomarker and facilitate precision medicine.

With that, I would like to thank you for your attention, and I'm happy to take questions and discuss further.

Andrea Miyahira: Thanks so much for sharing the study with us, Dr. Ding. Can it be determined what tumor or host, or tumor or microenvironment features the algorithm is seeing that pathologists are missing?

Cornelia Ding: Yeah. In this particular study, because we used an algorithm specifically designed to grade prostate cancer on the Gleason grading system, which is purely on the tumor architecture, we suspect there are probably very limited microenvironment features included in this algorithm. However, this is an ongoing effort now. We are trying to revisit these cases that are upgraded by AI but not by pathologists, and see what features are missing. I predict that we will probably see mostly information from the tumor itself. However, we have a different project also working on a more open-ended algorithm, not just based on Gleason grading, but on the entire slide evaluation. That could provide more information on prostate stroma, inflammatory cells, and their roles in disease progression.

Andrea Miyahira: Thank you. What are the next steps in clinical development of this algorithm, and how do you anticipate it being used in the clinic?

Cornelia Ding: Currently, I think AIRA MATRIX already has this algorithm embedded in many digital pathology systems that could help assist with Gleason grading. That could be helpful in saving pathologists' time. But furthermore, we are hoping that we can refine our pathology diagnosis by providing more precise information in managing active surveillance patients. The major limitation of the current study is that pathology grade reclassification does not necessarily mean the patient will have a bad outcome. Some patients might still have indolent disease, while others might have more rapid progression and even aggressive disease later. I'm hoping we can further associate the findings with patient outcomes, in addition to the pathology reclassification, which will provide more relevant information to the clinical team.

Andrea Miyahira: Okay. Thank you. What other uses for AI in pathology are being explored for patients with prostate cancer?

Cornelia Ding: In addition to Gleason grading, one of the recent exciting discoveries from other AI groups is providing predictive and prognostic biomarkers based on H&E images, such as responses to androgen deprivation therapy. I think that's a very exciting development, and it could be helpful in deciding what treatment or management strategy could be beneficial for each specific patient.

Andrea Miyahira: Okay. Well, thank you so much for coming on and sharing this with us today.

Cornelia Ding: Thank you.