Abstract
Automatic and accurate Gleason grading of histopathology tissue slides is crucial for prostate cancer diagnosis, treatment, and prognosis. Usually, histopathology tissue slides from different institutions show heterogeneous appearances because of different tissue preparation and staining procedures, thus the predictable model learned from one domain may not be applicable to a new domain directly. Here we propose to adopt unsupervised domain adaptation to transfer the discriminative knowledge obtained from the source domain to the target domain without requiring labeling of images at the target domain. The adaptation is achieved through adversarial training to find an invariant feature space along with the proposed Siamese architecture on the target domain to add a regularization that is appropriate for the whole-slide images. We validate the method on two prostate cancer datasets and obtain significant classification improvement of Gleason scores as compared with the baseline models.
Introduction
Gleason grading—the process of scoring the severity of prostate cancer from histopathology slides—is a cornerstone of diagnosis and treatment planning. While deep learning models, specifically Convolutional Neural Networks (CNNs), have shown great promise in automating this task, they suffer from a major "brittleness" problem: a model trained on images from one hospital often performs poorly on images from another.
This drop in performance is driven by domain shift. Different institutions utilize various microscope scanners, tissue preparation techniques, and staining procedures (such as H&E staining), leading to significant variations in image appearance. Traditionally, overcoming this shift requires pathologists to manually label thousands of new images for every new site—a process that is expensive, time-consuming, and prone to human error. This research proposes a method to adapt models to new sites without requiring any new manual labels.
Methodology
The proposed framework utilizes a two-stage unsupervised domain adaptation approach to bridge the gap between source and target domains.
1. Source Training and Adversarial Adaptation
Initially, a modified AlexNet is trained on a labeled "source" dataset using standard supervised learning to classify patches into low-grade (Gleason 6-7) or high-grade (Gleason 8-10). To adapt this model to a new "target" domain, the researchers employ a Generative Adversarial Network (GAN) philosophy. A discriminator network ($D$) attempts to distinguish whether a feature map originated from the source mapping ($M_s$) or the target mapping ($M_t$). Simultaneously, the target mapping network is trained to "fool" the discriminator, effectively forcing the model to learn features that are "domain-invariant."
2. Siamese Regularization
A key innovation of this paper is the Siamese architecture used for regularization. Since multiple patches are cropped from a single Whole-Slide Image (WSI), they should logically share the same Gleason grade. The Siamese network ensures that the model treats patches from the same slide consistently. This acts as a powerful unsupervised constraint, guiding the adaptation process toward features that are clinically relevant across the entire slide.
Experimental Setup and Datasets
The study utilized two primary data sources to validate the model's generalizability:
- TCGA (The Cancer Genome Atlas): A public dataset containing WSIs from 32 different institutions. A subset from the University of Pittsburgh (UP) served as one target domain.
- CINJ (Cancer Institute of New Jersey): A local, independent dataset used as a second target domain.
WSIs were preprocessed by cropping them into $2048 \times 2048$ patches, retaining only those with sufficient tissue area for analysis.
Key Results
The proposed method significantly outperformed baseline models that were trained on source data and applied directly to target data without adaptation.
- TCGA (w/o UP) → UP: Accuracy increased from a baseline of 54.3% to 77.1% using the full proposed method.
- TCGA → CINJ: Accuracy improved from 56.3% to 75.0%.
- State-of-the-Art Comparison: On the source domain, the modified AlexNet achieved 83.0% accuracy, outperforming previous benchmarks of 73.5%.
Ablation studies confirmed that while adversarial training alone improves performance, the addition of Siamese regularization provides a vital extra boost by leveraging the spatial context inherent in pathology slides.
Conclusion and Future Work
This work demonstrates that unlabeled data can be effectively used to align feature distributions across domains, overcoming the challenges of institutional heterogeneity in medical imaging. By introducing Siamese regularization, the authors leverage the biological relationship between patches from the same slide as a "free" supervision signal.
Future work will focus on moving beyond binary classification (low vs. high grade) to distinguish between specific Gleason scores (e.g., 3+4 vs. 4+3), which is critical for clinical treatment decisions. Additionally, testing the framework on larger and more diverse multi-institutional datasets will be necessary to ensure its robustness for clinical deployment.
Summary updated