Abstract
Prostate cancer (PCa) is one of the most common cancers in men around the world. The most accurate method to evaluate lesion levels of PCa is microscopic inspection of stained biopsy tissue and estimate the Gleason score of tissue microarray (TMA) image by expert pathologists. However, it is time-consuming for pathologists to identify the cellular and glandular patterns for Gleason grading in large TMA images. We used Gleason2019 Challenge dataset to build a convolutional neural network (CNN) model to segment TMA images to regions of different Gleason grades and predict the Gleason score according to the grading segmentation. We used a pre-trained model of prostate segmentation to increase the accuracy of the Gleason grade segmentation. The model achieved a mean Dice of 75.6% on the test cohort and ranked 4th in the Gleason2019 Challenge with a score of 0.778 combined of Cohen's kappa and the f1-score.
Introduction
Prostate cancer (PCa) remains a significant global health challenge. The definitive diagnostic tool for assessing PCa severity is the Gleason score, determined by pathologists through the microscopic examination of stained biopsy tissue. Pathologists assign a grade from 1 to 5 based on histological patterns, with the final score being the sum of the two most prevalent patterns.
Despite being the "gold standard," this manual grading process faces two critical hurdles:
- Efficiency: Manually identifying complex cellular patterns across large tissue microarray (TMA) images is labor-intensive and slow.
- Subjectivity: High inter-observer variability often leads to inconsistent grading between different experts, complicating clinical decision-making.
To address these issues, this research develops a Convolutional Neural Network (CNN) based on the U-Net architecture to automate the segmentation of TMA images into distinct Gleason grade regions.
Methods
The proposed pipeline utilizes a U-Net architecture, a standard for medical image segmentation, with several key enhancements to improve robustness and accuracy.
Handling Expert Variability
A core innovation of this work is the treatment of expert annotations. Rather than relying on a single "ground truth" label, the authors integrated annotations from six different pathologists. These annotations were encoded into separate channels and merged to generate a consensus probability map. This approach allows the model to learn from the distribution of expert opinions, effectively "smoothing" the ground truth and reducing the impact of individual outliers.
Transfer Learning and Preprocessing
The study explored the impact of cross-modality transfer learning. Two models were compared:
- Model 1: Initialized with random weights.
- Model 2: Pre-trained on a task involving the segmentation of prostate regions in 150 T2-weighted MRI images.
Surprisingly, pre-training on radiology data (MRI) improved performance on pathology images (histology), suggesting that low-level features related to prostate tissue structure are transferable across modalities.
For data management, images were down-sampled by a factor of 10 and normalized using Z-score standardization. Post-processing involved a 55×55 median filter to eliminate noise before up-sampling to the original resolution.
Datasets
The model was trained and evaluated using the Gleason2019 Challenge dataset, comprising 244 TMA images. The data was partitioned into training (188), validation (33), and testing (23) sets. Pixels were classified into categories including benign, Gleason grades 3, 4, 5, and "ignored" regions.
Key Results and Insights
The pre-trained model (Model 2) achieved a 4th-place finish in the Gleason2019 Challenge, demonstrating high reliability.
- Performance Metrics: The model reached a mean Dice coefficient of 75.6% and a combined challenge score (Cohen’s kappa and F1-score) of 0.778.
- Class-Specific Accuracy: Performance was strongest for benign tissue (88.32%) and Grade 3 cancer (66.57%). However, accuracy declined for more aggressive stages, reaching 46.93% for Grade 4 and 22.90% for Grade 5.
- Reliability: 95% of test cases achieved a Dice score above 0.6, indicating consistent performance across diverse samples.
The results highlight that while deep learning is highly effective at identifying benign and low-grade patterns, the chaotic cellular structures of high-grade (4 and 5) cancers remain a significant challenge, likely due to their inherent complexity and the relative scarcity of training data for these severe cases.
Conclusion
This study demonstrates the potential of deep learning to standardize and accelerate Gleason grading. By leveraging multi-expert consensus and cross-modality transfer learning, the proposed model provides a robust framework for automated PCa diagnosis. Future work will focus on addressing data imbalance for high-grade cases and exploring advanced architectures like Generative Adversarial Networks (GANs) to further refine feature extraction from limited datasets.
Summary updated