Andrea Miyahira: Hello, everybody. I am Andrea Miyahira on the Prostate Most cancers Basis. Right here with me is Dr. Cornelia Ding, an assistant professor at UCSF. She’s going to talk about her current paper, “Predicting Prostate Most cancers Grade Reclassification on Energetic Surveillance utilizing a Deep Studying-based Grading Algorithm,” that was printed in JNCI. Dr. Ding, thanks a lot for becoming a member of us right now.
Cornelia Ding: Thanks for having me. I am very excited to share this work from after I was a pathology fellow at Johns Hopkins College, working with Dr. De Marzo and Dr. Lotan, about predicting prostate most cancers grade reclassification utilizing AI.
That is collaborative work from many individuals with totally different specialties. So, after all, pathologists are the foremost drivers of this research, but additionally our urologists, oncologists, and our collaborator in India, AIRA MATRIX, offering the AI algorithm evaluating histology slides, are concerned on this work.
Right here is the disclosure. The research on this presentation is supported by a grant from the Prostate Most cancers Basis, sponsored by AIRA MATRIX. And, importantly, AIRA MATRIX didn’t help with knowledge evaluation or writing of this research.
Scientific administration for localized prostate most cancers is strongly pushed by pathology. By interpretation of prostate biopsies, we pathologists present a analysis of prostate most cancers in addition to grade group, which relies on analysis of tumor structure utilizing the Gleason grading system underneath the microscope.
The tumor structure has been proven to be strongly related to prostate most cancers affected person outcomes. As proven right here, the Y-axis is biochemical recurrence-free survival, and the X-axis is time. You possibly can see sufferers with grade group 1 illness have indolent illness development over time. In order that they normally don’t profit from speedy definitive remedies resembling radical prostatectomy. Many sufferers these days with grade group 1 illness keep on lively surveillance.
Nonetheless, a minority of those sufferers with low-grade illness could have illness development over time. So our fundamental query is, can we refine our pathology analysis of prostate biopsies to raised determine sufferers who could have illness development? And might AI, synthetic intelligence, assist us right here?
With collaboration from AIRA MATRIX, they developed a deep learning-based algorithm referred to as AIRAProstate for tumor detection and Gleason grading utilizing hematoxylin and eosin stained pathology whole-slide pictures.
On this specific research, we evaluated two totally different cohorts at Johns Hopkins College. The primary cohort is a case-control design involving 138 sufferers with grade group 1 illness on analysis. Sixty-four sufferers later had grade reclassification, so they’re the instances. Seventy-four sufferers stayed on lively surveillance with none grade reclassification for greater than eight years. So they’re the controls.
One H&E slide per affected person is regraded by each up to date uropathologists and the AI algorithm AIRAProstate.
On the appropriate aspect, you’ll be able to see a head-to-head comparability from up to date pathologist grading versus AIRAProstate. Pathologists and AIRAProstate largely agree upon the detection of most cancers on every slide with 95% settlement. Nonetheless, as a result of these sufferers have been enrolled a few years in the past and the grade system has been up to date, the up to date pathologists don’t essentially agree with the preliminary analysis of grade group 1.
A few of them are upgraded into grade group 2 or larger. Nonetheless, a proportion aren’t considerably totally different. In distinction, upgrading by AIRAProstate is totally different between case and management, the place 33% of instances are upgraded, whereas 8% of controls are upgraded. So you’ll be able to see the chances ratio is 3.3 with P 0.04.
A special cohort can also be evaluated. It is a little totally different as a result of all of the sufferers had MRI previous to the preliminary prostate biopsy, and nearly all of the sufferers have been enrolled after 2014. There are 169 sufferers with grade group 1 illness on this cohort; 94 of them stayed on lively surveillance with out grade reclassification throughout follow-up. Nonetheless, 75 sufferers had later grade reclassification. Thirty-five of those sufferers underwent radical prostatectomy. H&E slides with cancers are evaluated by the AI algorithm AIRAProstate. You possibly can see right here substantial instances additionally get upgraded by the AI algorithm. Eighteen p.c have been upgraded by AI. Nonetheless, 40% of the instances that had later grade reclassification have been additionally upgraded. And these are totally different by hazard ratio 1.7, P equals 0.3.
In abstract, that is the primary research within the literature evaluating a deep learning-based prostate most cancers grading algorithm to medical outcomes. We in contrast two totally different lively surveillance cohorts at Johns Hopkins, each displaying that the upgrading of preliminary biopsy by AI is related to speedy or excessive grade reclassification, which predicts later illness development.
So a take-home message from this research is that we’ve got been utilizing histopathology as biomarkers and in addition because the gold normal of analysis. This has offered necessary info that drives medical choices, particularly in managing localized prostate most cancers. These days, with assist from synthetic intelligence in pathology, we may doubtlessly refine this biomarker and facilitate precision drugs.
With that, I wish to thanks to your consideration, and I am glad to take questions and talk about additional.
Andrea Miyahira: Thanks a lot for sharing the research with us, Dr. Ding. Can it’s decided what tumor or host, or tumor or microenvironment options the algorithm is seeing that pathologists are lacking?
Cornelia Ding: Yeah. On this specific research, as a result of we used an algorithm particularly designed to grade prostate most cancers on the Gleason grading system, which is solely on the tumor structure, we suspect there are in all probability very restricted microenvironment options included on this algorithm. Nonetheless, that is an ongoing effort now. We are attempting to revisit these instances which might be upgraded by AI however not by pathologists, and see what options are lacking. I predict that we’ll in all probability see principally info from the tumor itself. Nonetheless, we’ve got a distinct challenge additionally engaged on a extra open-ended algorithm, not simply based mostly on Gleason grading, however on your entire slide analysis. That would present extra info on prostate stroma, inflammatory cells, and their roles in illness development.
Andrea Miyahira: Thanks. What are the subsequent steps in medical improvement of this algorithm, and the way do you anticipate it getting used within the clinic?
Cornelia Ding: At present, I feel AIRA MATRIX already has this algorithm embedded in lots of digital pathology methods that would assist help with Gleason grading. That might be useful in saving pathologists’ time. However moreover, we hope that we will refine our pathology analysis by offering extra exact info in managing lively surveillance sufferers. The main limitation of the present research is that pathology grade reclassification doesn’t essentially imply the affected person may have a foul consequence. Some sufferers may nonetheless have indolent illness, whereas others may need extra speedy development and even aggressive illness later. I am hoping we will additional affiliate the findings with affected person outcomes, along with the pathology reclassification, which is able to present extra related info to the medical workforce.
Andrea Miyahira: Okay. Thanks. What different makes use of for AI in pathology are being explored for sufferers with prostate most cancers?
Cornelia Ding: Along with Gleason grading, one of many current thrilling discoveries from different AI teams is offering predictive and prognostic biomarkers based mostly on H&E pictures, resembling responses to androgen deprivation remedy. I feel that is a really thrilling improvement, and it might be useful in deciding what therapy or administration technique might be useful for every particular affected person.
Andrea Miyahira: Okay. Nicely, thanks a lot for approaching and sharing this with us right now.
Cornelia Ding: Thanks.
