Prostate cancer is one of the most common cancers in men worldwide. The gold standard for diagnosis remains the microscopic examination of tissue biopsies stained with hematoxylin and eosin (H&E), which allows pathologists to detect tumor cells and assess their aggressiveness using Gleason grading.
Gleason grading ranks cancer patterns on a scale of 1 to 5 based on gland structure. When two patterns are present, the scores of the most common and the most aggressive patterns are added together. The resulting combined score correlates with prognosis. An updated system called the ISUP grading system groups Gleason scores into five categories, from indolent (group 1) to aggressive (group 5).
A key technological shift enabling AI pathology is the digitization of glass slides into whole slide images (WSIs) using high-resolution scanners. Once in digital form, slides can be fed into deep learning algorithms that learn to identify patterns associated with cancer, grade, or prognosis.
This systematic review covered 77 studies published through 2022, cataloguing deep learning applications across three major task categories: pre-processing (image quality and staining normalization), diagnosis (cancer detection and Gleason grading), and prediction (outcomes and genomic features).
Two main types of algorithms are used in this field. Segmentation algorithms precisely outline regions of interest at the pixel level, such as identifying malignant versus benign glands. The most popular architecture for segmentation is U-Net. Classification algorithms assign labels such as cancer grade to tiles, slides, or patients.
Because WSIs are enormous -- far too large to process as a single image -- they must be divided into thousands of smaller tiles. Each tile is then encoded into numerical features using convolutional neural networks (CNNs). Popular CNN architectures used in these studies include ResNet, InceptionV3, VGG, DenseNet, and MobileNet, often adapted through transfer learning from models pre-trained on large natural image datasets.
Multiple Instance Learning (MIL) is a particularly powerful approach that assigns a probability score to each tile and uses the most informative tiles to make a slide-level or patient-level diagnosis. This method addresses the challenge of different WSI sizes and also provides explainability by highlighting which regions of the slide drove the prediction.
All studies were evaluated for risk of bias using an adapted version of the QUADAS-2 tool, assessing criteria including whether ground truth was defined by multiple experts, whether datasets were publicly available, and whether external validation cohorts were used.
Cancer detection was the most heavily studied application, with 21 studies focused exclusively on detecting malignant tissue and another 19 combining detection with Gleason grading. Many algorithms achieved AUC values above 0.95, with the best reaching 0.99 on internal validation cohorts.
The largest and most influential study, by Campanella et al., trained a Multiple Instance Learning model on over 12,000 slides and achieved an AUC of 0.99 internally and 0.93 on an external cohort of over 12,000 additional slides, demonstrating genuine generalizability across institutions.
The PaigeProstate algorithm, derived from this work, was approved by the FDA in 2021 for clinical use. Multiple external validation studies confirmed its high performance, with sensitivity of at least 0.94 and specificity of at least 0.93. Studies also showed it improved pathologist performance when used as a second-read tool, increasing sensitivity from 0.74 to 0.90 in one cohort.
Beyond simple cancer detection, algorithms have been developed for more specific patterns. Detection of cribriform patterns, an aggressive gland structure, reached accuracies of 0.88. Detection of perineural invasion, where cancer grows along nerve sheaths, achieved an AUC of 0.95. These subtypes are clinically important but difficult for pathologists to detect consistently.
Gleason grading is a harder task than cancer detection because it requires distinguishing between cancer patterns that share subtle structural similarities. Performance is often measured with quadratic Cohen Kappa, a statistic that ranges from 0 (random agreement) to 1 (perfect agreement), with pathologists typically scoring around 0.65 to 0.75 against a reference standard.
The PANDA challenge (Prostate cANcer graDe Assessment) provided a landmark dataset of approximately 12,000 biopsy slides from 3,500 patients across two institutions, with external validation from two additional sites. The best algorithms in this challenge reached a quadratic Kappa of 0.85 on external validation, matching or exceeding typical pathologist performance.
Commercially available tools for Gleason grading include DeepDx (quadratic kappa 0.90 internally, 0.90 on external validation by an independent group) and IBEX Galen Prostate (AUC 0.99 for cancer detection, AUC 0.94 for low vs. high grade differentiation). IBEX became the first such algorithm to receive CE marking under the EU In Vitro Diagnostic Medical Devices Regulation in early 2023.
Pixel-level segmentation of individual Gleason patterns is more challenging, with Intersection over Union (IoU) scores ranging from 0.48 to 0.79 depending on the study. Adding epithelium detection as a prior step improved mean IoU by 0.07 in one study, suggesting that preprocessing and tissue segmentation meaningfully boost grading accuracy.
The most ambitious and exploratory application is using biopsy slide images to predict future clinical events -- such as cancer recurrence, metastasis, or death -- without relying on any additional clinical information. This field is called pathomics-based prognosis.
Several studies demonstrated that deep learning models could stratify patients by recurrence risk with AUC values around 0.78 to 0.81 and hazard ratios (HR) between 4.8 and 7.1, meaning patients in the high-risk group had up to 7 times the recurrence rate of low-risk patients. One model predicted distant metastasis at 5 years with an AUC of 0.84 by combining image features with clinical data.
A particularly novel application involved identifying cancer in benign-appearing biopsy cores from patients with cancer elsewhere in the prostate. One model achieved AUC 0.74 on this task, suggesting that AI can detect subtle field effects -- microscopic changes surrounding a tumor that are invisible to the human eye.
Five studies explored whether slide images could predict genomic mutations directly, including SPOP mutations (AUC 0.86) and ERG gene rearrangement status (AUC 0.80 to 0.85). These findings suggest that tumor morphology encodes molecular information, potentially enabling genomic risk stratification without expensive DNA testing.
Risk of bias analysis revealed several recurring problems across the 77 studies. Only 24 of 77 papers included an external validation cohort -- a critical requirement to prove that a model generalizes beyond the hospital where it was trained. Without external validation, strong published results may reflect overfitting to a specific scanner, staining protocol, or patient population.
Only 32 of 77 studies had ground truth annotations defined by multiple expert pathologists. This is especially important for Gleason grading, where inter-observer variability among pathologists is itself around 0.65 to 0.75 Kappa. Training an algorithm on a single pathologist's annotations may teach the model to replicate one person's biases.
Most studies (50 of 77) did not apply staining normalization, even though slides from different hospitals or scanners show significant color variation that degrades model performance. Similarly, only 41 of 77 addressed class imbalance -- the fact that cancer tissue represents a small fraction of most biopsy slides -- which can bias models to favor predictions of the majority class.
Only 15 of 77 studies publicly shared their code, and only 21 used publicly available datasets. This lack of reproducibility makes it difficult for other researchers to validate results or build on prior work, slowing field-wide progress. The authors call for standardized reporting, public code sharing, and broader use of challenge datasets like PANDA.
Several AI tools for prostate cancer pathology are now commercially available and CE or FDA cleared: PaigeProstate (Paige), Galen Prostate (Ibex), DeepDx (DeepBio), and the Aiforia platform. These can operate as a first-read screening tool -- flagging suspicious slides -- or as a second-read quality check after the pathologist has reviewed a case.
Implementing AI in clinical labs requires more than good algorithms. It demands a digital pathology workflow including slide scanners, a laboratory information system, and an image management system with integrated AI. These infrastructure requirements limit deployment to larger academic or commercial labs and remain a significant barrier in resource-limited settings.
For prognosis, the review argues that the biggest opportunity lies in multimodal models that combine pathology images with MRI scans, blood biomarkers, genomic data, and clinical variables. Cancer prognosis is inherently multifactorial, and no single data modality is likely to capture all relevant information. How to optimally combine these different data types remains an open research question.
Explainability is identified as a key requirement for clinical adoption. Models that generate heatmaps highlighting which regions of the slide drove the prediction are more likely to be trusted and adopted by pathologists, and 40 of 77 studies attempted some form of explainability. Moving from opaque black-box models to transparent, interpretable systems is a priority for the field.