Prostate cancer (PCa) is the leading cause of cancer-related death in males, with over 1.4 million new cases diagnosed worldwide each year. Accurate diagnosis and grading are the foundation of all treatment decisions for this disease.
After a suspicious PSA (prostate-specific antigen) blood test, patients typically undergo a biopsy where 12 or more tissue samples are taken with a needle. Each sample must be examined under a microscope by a pathologist, who determines whether cancer is present and assigns a Gleason grade group (GG) -- a scoring system that reflects how aggressive the cancer looks.
With over 15 million biopsy samples processed worldwide each year, the workload on pathologists is enormous. This is compounded by a global shortage of pathologists and known variability in grading between different pathologists -- a phenomenon known as inter-observer inconsistency -- which can lead to some aggressive cancers being under-treated and some non-aggressive cancers being over-treated.
These challenges have driven significant interest in using artificial intelligence (AI) to standardize, accelerate, and improve the accuracy of prostate cancer pathology review.
Modern pathology AI relies on whole-slide imaging (WSI) -- digitizing glass slides into high-resolution images that can be analyzed by computer. This digitization is the enabling technology that makes AI-powered pathology possible.
The AI models used are primarily deep learning (DL) systems based on convolutional neural networks (CNNs). Unlike older machine learning approaches that required experts to manually define what features to look for, deep learning automatically discovers relevant image patterns from large numbers of training examples.
Because whole-slide images are extremely large, they are divided into smaller patches (tiles) which are fed into the model. Several training strategies are used -- from fully supervised learning (where pathologists annotate every gland) to weakly supervised learning (where only the overall slide diagnosis is used as a label), which dramatically reduces the annotation burden.
Before clinical deployment, AI models must be validated on independent datasets from different institutions to ensure they can generalize beyond their training data, and compared directly against human pathologists to establish their real-world accuracy.
AI has achieved impressive performance at the core task of detecting cancer in biopsy slides. The FDA-approved Paige Prostate system, trained on over 12,000 whole-slide images, achieved an AUC of 0.99 on its test set. In a user study, pathologist sensitivity for detecting cancer improved from 74% to 90% when working with AI assistance.
Automated Gleason grading -- determining how aggressive a cancer looks microscopically -- is a harder task, but AI has shown it can match expert uropathologist performance. The PANDA Challenge, a major international competition, produced algorithms trained on 10,616 biopsy samples that achieved concordance scores of 0.862-0.868 with expert graders.
AI grading is particularly useful at the clinically critical boundary between low-grade (Grade Group 1) -- suitable for active surveillance rather than immediate treatment -- and higher-grade cancers requiring intervention. Distinguishing these groups accurately prevents unnecessary surgery and radiation in patients who do not need them.
Some AI systems have proposed fine-grained grading concepts that go beyond the existing Gleason system, subdividing patterns into intermediate categories (such as GP 3.5 and GP 4.5) that may provide more accurate risk stratification than the current five-tier system.
Beyond grading, AI is showing promise in predicting patient outcomes directly from tissue images. A model developed to quantify specific cancer patterns achieved a C-index of 0.87 for predicting prostate cancer-specific mortality -- substantially outperforming the conventional Gleason Grade Group system (C-index 0.75).
Biochemical recurrence (BCR) -- a rise in PSA after treatment signaling cancer return -- is an important early outcome measure. Multiple AI models have successfully predicted BCR directly from post-surgery tissue images, outperforming standard clinical indicators. Combining AI predictions with Gleason grading produced the best performance in several studies.
AI can also detect molecular alterations from tissue images alone -- without requiring expensive genetic sequencing. Models have been trained to predict common prostate cancer mutations including TMPRSS2-ERG gene fusions, PTEN deletions, and SPOP mutations from standard H&E-stained slides, with AUC values ranging from 0.76 to 0.84.
This ability to predict molecular characteristics from routine tissue slides -- which are available for virtually all patients -- could make molecular subtyping accessible in settings where genetic sequencing is unavailable or cost-prohibitive, representing a significant advance for resource-limited healthcare environments.
A recurring theme in this review is that AI works best as a collaborative partner with human pathologists rather than as a replacement. Multiple studies show that the AI-plus-pathologist combination outperforms either alone in accuracy and consistency.
As a first-pass screener, AI can pre-screen all biopsy slides and identify those that appear benign, allowing pathologists to focus their time on slides with suspected cancer. Since a substantial proportion of biopsies are benign, this could dramatically reduce workload without compromising diagnostic safety.
As a second reader, AI can flag cases where its assessment differs from the pathologist's -- helping catch small cancers that might be missed, preventing under-grading of aggressive tumors, and providing an objective check on the human diagnosis. In one study, AI assistance increased detection sensitivity for small cancers (less than 0.6 mm) from 46% to 83%.
AI also enables practical workflow improvements such as automatically identifying which cases require additional immunohistochemistry (IHC) staining before pathologist review, reducing turnaround time and improving laboratory efficiency. It can also provide quality control on slide scanning and prioritize difficult cases for expert review.
Generalization is a major hurdle. AI models trained at one institution often perform worse when applied to data from other centers, due to differences in tissue staining, scanning equipment, and patient populations. Strategies to address this include large diverse training sets, data augmentation, and image style normalization techniques.
The quality of ground truth labels -- the expert annotations used to train AI -- is a fundamental challenge. Since different pathologists sometimes disagree on Gleason grades, AI trained on one pathologist's annotations may inherit that individual's biases. Best practices now involve using consensus annotations from multiple experts.
Most current AI models are black boxes -- they produce predictions but cannot easily explain why. This limits pathologist trust and can hide cases where the AI is using unexpected or misleading features. Techniques like gradient activation mapping (Grad-CAM) and automatic concept explanations are being used to improve interpretability.
Broader adoption also faces infrastructure barriers -- most pathology labs worldwide are not yet fully digitized. Scalable deployment will require cost-effective whole-slide scanners and interim solutions like AI-augmented microscopes that can add AI capability to conventional lab equipment without full digitization.
Recent advances in deep learning have brought AI pathology to a level of performance that is comparable to -- and in some cases better than -- expert pathologists for specific tasks. The regulatory approval of the first FDA-cleared AI pathology tool for prostate cancer marks a pivotal transition from research to clinical reality.
Future AI systems are expected to go beyond replicating current pathological assessments and instead discover entirely new histological features and biomarkers that human pathologists have not identified -- potentially uncovering new mechanisms of cancer progression.
Combining AI pathology with spatial omics technologies -- which map genetic information to specific locations within a tissue -- could provide unprecedented insights into tumor heterogeneity, showing how different cancer subpopulations coexist within a single tumor and interact with the immune system.
Ensuring these tools work equitably across diverse patient populations is critical. Current large AI studies have been predominantly based on Western populations, and biological differences in prostate cancer across racial groups -- such as higher rates of FOXA1 mutations in Asian patients -- mean that global validation studies are urgently needed.