Artificial intelligence unravels interpretable malignancy grades of prostate cancer on histology images

Prostate cancer (PCa) is a prevalent malignancy with variable aggressiveness and prognosis. Malignancy grading is crucial for treatment decisions, currently relying on the modified Gleason grading system. However, this system suffers from significant interobserver variability, even among experienced pathologists. While deep learning shows promise in improving Gleason grading consistency, existing AI approaches inherit the limitations of the current grading system. This study aims to overcome these limitations by developing a novel AI-based grading system for PCa that predicts recurrence and survival outcomes using long-term PCa prognosis data from a large, multi-institutional international dataset, rather than directly relying on the subjective interpretation of microscopic histology. This novel approach leverages the tissue microarray (TMA) framework of the Canadian Prostate Cancer Biomarker Network (CPCBN), ensuring the collection of representative PCa samples from radical prostatectomy (RP) specimens. The objective is to create a calibrable and interpretable algorithm for predicting PCa outcomes that can be integrated into existing prognostic tools.

Existing literature highlights the potential of artificial intelligence (AI) in improving the consistency of Gleason grading for prostate cancer. Studies have demonstrated that AI models can achieve accuracy comparable to that of expert pathologists. However, these studies often rely on the current Gleason grading system as the ground truth, inheriting its inherent limitations in terms of interobserver variability and potential biases. The current Gleason grading system, despite revisions, continues to face challenges in reproducibility due to interobserver variability in grading and quantification, leading to inconsistencies. Therefore, this study proposes a novel approach to overcome these limitations by focusing on long-term prognosis instead of relying solely on microscopic histology.

The study utilized data from multiple independent cohorts: a development cohort (n=600) and three external validation cohorts (CPCBN, n=890; PROCURE, n=287; PLCO, n=861). Tissue microarray (TMA) slides were scanned, and tissue regions extracted using QuPath. These images were tiled into patches labeled by biochemical recurrence (BCR) status. A novel AI model was developed to predict BCR using histology images as input and a confidence score as output. The model's architecture was optimized using neural architecture search and grid search. Various deep learning architectures (e.g., ResNet, VGG, EfficientNet) were compared, resulting in a superior model with fewer parameters and better performance. The model's performance was evaluated using metrics such as c-index, AUROC, sensitivity, specificity, precision, and recall. Calibration plots assessed the model's accuracy in predicting 5-year and 10-year BCR. Univariate and multivariate Cox regression analyses were performed to assess the prognostic value of the novel risk classification system for biochemical recurrence and cancer-specific survival, comparing it to the traditional Gleason grade group (GG). A CHAID analysis was used to define risk groups based on the BCR scores. The interpretability of the model was evaluated through feature analysis (SHAP and LIME) and assessment of concordance between pathologist interpretations and the AI-generated risk classifications. Kaplan-Meier curves were used to visualize survival probabilities across risk groups. The association between the novel risk score and the development of castration-resistant prostate cancer (CRPC) was investigated using correlation analysis and multivariate Cox regression.

The AI model developed a novel BCR prediction system that outperformed existing models in terms of prediction accuracy and model parsimony. The model demonstrated good calibration and high predictive performance (c-index = 0.682 in the first external validation set). The novel risk classification system, derived from the AI model's BCR scores, showed superior prognostic capabilities compared to the traditional Gleason grade grouping (GG) for both biochemical recurrence and cancer-specific survival in all three external validation cohorts. The four risk groups defined by the CHAID analysis (low, intermediate, high-intermediate, high) had significantly different survival probabilities. The novel system also demonstrated independent prognostic value for the development of CRPC in men experiencing biochemical recurrence. Five expert pathologists showed strong agreement with the AI-generated risk classifications, indicating the system's interpretability and clinical relevance. Analysis of representative features revealed a clear histopathological gradient across the risk groups, further supporting the system's biological validity.

This study successfully demonstrated the development and validation of a novel AI-based grading system for PCa that is superior to the current Gleason grading system. The AI model's ability to predict biochemical recurrence, cancer-specific survival, and CRPC risk, along with its high degree of interpretability, addresses the limitations of the current standard grading system. The strong agreement between the AI-generated risk classifications and the independent assessments of experienced pathologists underscores the clinical utility of this novel approach. The interpretability of the model is a significant advantage, facilitating communication and collaboration between AI and human experts in clinical decision-making. The results highlight the potential for AI to improve risk stratification and personalized treatment planning for PCa.

This study presents a novel AI-driven grading system for prostate cancer that demonstrates superior performance and interpretability compared to the existing Gleason score system. The system's ability to predict biochemical recurrence, cancer-specific survival, and CRPC risk provides valuable tools for improved risk stratification and personalized treatment. Future research could focus on integrating this system into clinical workflows and exploring its potential for use in other cancer types. Further investigation into the biological underpinnings of the identified features could further refine and enhance the system's accuracy and clinical utility.

While the study demonstrates robust performance across multiple cohorts, limitations exist. The availability of data limited open access. Further studies using larger, more diverse datasets are warranted to confirm the generalizability of the findings. The study's reliance on radical prostatectomy samples may limit the applicability of the findings to other treatment approaches.