Introduction
Prostate cancer (PCa) is one of the leading causes of cancer-related deaths in American men. PCa risk diagnostics and stratification involves multiple steps including prostate-specific antigen (PSA) test, digital rectal exam, biopsy, and Gleason grading. Prostate biopsy which includes histopathology image analysis, is an integral part of the diagnosis as it helps make an accurate diagnosis and guide clinical management. The whole slide images (WSIs) from histopathology analysis provide detailed morphological data, but their large size and complexity make it difficult to extract high-quality features for predictive modeling. This study introduces an automated pipeline to extract high-quality histopathology data from WSIs and develop models to predict prostate cancer risk. Using advanced image processing and nuclei scoring, our approach aims to enhance the accuracy of risk assessment, support personalized diagnostics, and improve patient outcomes.
Methodology
We developed a novel pipeline using the Histolab Python library to process WSIs from a cohort of 50 PCa patients, comprising 25 high-risk and 25 low-risk prostate cancer cases, based on clinical annotations. The pipeline employs the NucleiScorer() function to rank image regions by nuclei density and morphological features (shape and appearance), extracting the 15 highest-scoring 512×512 pixel tiles per WSI. Increased nuclear density and size along with irregular shapes in a histopathology tissue sample, are frequently associated with higher cancer aggressiveness. Histolab assigns a score to each WSI tile based on nuclear density and morphological attributes, where higher scores indicate greater relevance (i.e., cancerous regions), and lower scores correspond to lower relevance (i.e., benign or less cancerous regions). These tiles capture essential histopathological characteristics and were pre-processed to ensure consistency in staining and resolution. The extracted tiles were used to train machine learning models to predict prostate cancer risk. Feature analysis identified key histopathological patterns associated with risk levels, improving the interpretability of the model.
Results
The developed pipeline successfully extracted high-quality tiles enriched with nuclear features, as confirmed by visual inspection and quantitative scoring. The results of the extracted tiles are categorized into per-patient mean highest scores (0.61), per-patient mean lowest scores (0.44), high risk patients mean highest scores (0.65), low risk patients mean highest scores (0.57), and mean aggregated highest score all patients (0.61). The resulting models showed promising performance in distinguishing high-risk patients from low-risk patients, with nuclear density and morphology emerging as significant predictors. The automated tile selection process reduced computational demands while maintaining data quality, enabling scalable model development. The ongoing work focuses on expanding the dataset and integrating additional clinical data to further enhance predictive accuracy.
Conclusion
This developed pipeline demonstrates the potential of automated WSI tiling and machine learning for the prediction of prostate cancer risk. By prioritizing high- quality histopathological data through nuclei scoring, our approach provides a foundation for accurate and efficient risk stratification.This framework enables personalized prostate cancer care via improved models and multimodal data.
