<p>Artificial intelligence-digital pathology (AI-DP) holds great promise in enhancing the understanding of prostate cancer across the disease continuum. However, further prospective studies are needed to validate multimodal AI (MMAI) models as both a potential prognostic and predictive tool. A presentation by Philip Sutera, MD, at the <strong>2025 ASTRO Annual Meeting</strong> included information on the potential use of AI-DP in patients with oligometastatic castration-sensitive prostate cancer.</p> <p><br></p> <p><em>Following this presentation, featured expert Daniel E. Spratt, MD, was interviewed by </em>Conference Reporter <em>Editor-in-Chief Tom Iarocci, MD. Clinical perspectives from Dr Spratt on these findings are presented here.</em></p>

Pathology has evolved from evaluating histopathology using traditional microscopy to using digitized images that can be analyzed with computational models and machine learning. AI-DP has evolved to include multiple variables, allowing for its ability to assist in cancer grading, prognosticate outcomes, and potentially predict treatment response. AI-DP holds great promise in the management of prostate cancer, from screening and diagnosis to monitoring and survivorship.

<br>

MMAI classifiers such as the ArteraAI Prostate Test (Artera) can incorporate multiple forms of data, including DP, clinical features (eg, prostate-specific antigen and T stage), and the patient’s age. I participated in some of the development of and validation work on the ArteraAI Prostate Test, and we have applied it retrospectively to multiple clinical trials. The data have shown that MMAI can be prognostic, meaning that it can risk stratify patients and help determine which individuals have a lower or higher risk of developing metastatic disease.

<br>

The ArteraAI Prostate Test also shows promise as a predictive model. In the NRG Oncology/Radiation Therapy Oncology Group 9408 validation cohort, the ArteraAI Prostate Test was able to identify those with a predominantly intermediate risk for prostate cancer who were likely to benefit from short-term ADT use. I think that the validation was robust; however, because this was not a prospective randomized clinical trial, having another validation study would be beneficial. If we had a second study that successfully predicted those who would benefit from ADT using a differential trial population, that would give me more confidence.

<br>

One challenge with MMAI models is understanding what they are looking at and how much of the DP is being considered in the model output vs clinical features. Currently, as Dr Sutera discussed during his presentation at ASTRO 2025, this is a “black box” in our understanding of MMAI that needs to be opened, and clinicians may question whether they can trust it. I care more about the robustness of the MMAI model validation than I do about understanding exactly what it is looking at on the pathology side.

<br>

There are also studies investigating the use of DP algorithms for predicting mutational profiles (eg, TP53 mutations) across many different cancer types, although the outcomes have been variable. In prostate cancer, there is an interest in using algorithms to identify PAM50 subtypes. In a recent study, the model area under the curve was 0.78 for a model trained on digitized prostate biopsy slides, meaning that it is not just guessing but also that it is probably not accurate enough. I think that there will be limitations to how accurately an algorithm can predict a phenotype from a given genotype. Perhaps future models can incorporate additional features, such as magnetic resonance imaging scans or other clinical data. For now, I am not confident that you can predict specific prostate cancer subtypes using AI, but I think that AI can identify other useful features.

<br>

Further along in the disease continuum, MMAI has also shown promise in predicting prognosis in more advanced disease. Studies have shown how highly prognostic MMAI tools can be just from assessing tissue from the primary tumor biopsy—even years later. During his talk at ASTRO 2025, Dr Sutera also discussed a recent study evaluating a DP-based MMAI biomarker model in oligometastatic castration-sensitive prostate cancer. This study looked at 2 clinical trials: STOMP and ORIOLE. Although these trials were small, patients underwent exhaustive molecular testing that included circulating biomarkers and DP. This study showed that patients with higher MMAI scores derived the most benefit from metastasis-directed therapy, whereas it was harder to see those benefits in patients with lower MMAI scores.

<br>

Explainability studies will be critical for confirming the applicability of AI models. For simpler DP models, this could mean highlighting exactly which features on the DP slide are being evaluated. However, these types of studies can be quite challenging for MMAI models that incorporate more than just DP, such as the ArteraAI Prostate Test.

<br>

Looking forward, I think that incorporating additional model inputs, such as single-cell sequencing and/or spatial transcriptomics, can aid in detecting currently undetectable features on DP. I also see DP being used even further upstream to help detect cancer on a biopsy and to quantitively grade the cancer. The hope is that, eventually, the diagnostic, prognostic, and predictive functions of MMAI tools can be consolidated into a single robust package for physicians to use from diagnosis all the way through treatment response.