Source-linked AI summary
Adversarial Domain Adaptation for Classification of Prostate Histopathology Whole-Slide Images
Jian Ren, Ilker Hacihaliloglu, Eric A. Singer, David J. Foran, Xin Qi
TL;DR
Institutional differences in prostate histopathology WSIs can undermine Gleason classifiers trained on one domain when applied to another. The paper uses adversarial unsupervised domain adaptation with a target-domain Siamese regularizer, achieving statistically significant classification improvement over a baseline without additional target annotations.
Problem
Different scanners, staining procedures, and tissue distributions can make supervised Gleason classification ineffective on a new institution’s unlabeled dataset.
Source: §1 · pp. 1–2Method
The method aligns source and target feature distributions adversarially and uses a Siamese target network to regularize patches from the same WSI.
Source: §2 · pp. 3–4Results
Statistically significant classification improvement was achieved over the baseline for TCGA (w/o UP) →UP, with McNemar Test p-value 0.039.
Source: §3 · p. 6Takeaways & Limitations
Abstract
from arXiv · showhide
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.
1 Introduction
Institutional differences in scanners, staining, and tissue appearance can make supervised Gleason classification fail when transferred to another prostate dataset. The paper addresses this with unsupervised domain adaptation using adversarial alignment and a target-domain Siamese constraint.
1 Introduction
- Different institutions produce heterogeneous WSI appearances, making source-domain Gleason classifiers ineffective on target datasets.The variation is attributed to differences in microscope scanners, staining procedures, and glandular-region distributions.Source: §1 · pp. 1–2
- Unsupervised adaptation transfers discriminative knowledge from an annotated source domain to an unlabeled target domain without additional target labeling.The approach aligns source and target image distributions while assuming the same number of high-level classes.Source: §1 · p. 2
- Adversarial training minimizes feature-space distribution discrepancy to discover domain-invariant representations.The method adopts a GAN loss for the adaptation.Source: §1 · p. 2, Figure 1 · p. 2
- A target-domain Siamese architecture regularizes cropped patches from the same WSI, complementing adversarial adaptation.This constraint is designed for whole-slide images and their patch structure.Source: §1 · p. 2, Figure 1 · p. 2
- The method classifies newly given prostate datasets into low and high Gleason grades through unsupervised learning and is validated on public and newly collected datasets.The authors report significantly improved Gleason-score classification accuracy compared with a baseline model.Source: §1 · p. 2
2 Method
The method first learns supervised discriminative features from labeled source WSIs, then adapts a target network using adversarial feature alignment and Siamese patch regularization. Target-domain predictions are aggregated from cropped patches to obtain WSI-level Gleason grades.
Learning at source domain
- The source network uses cropped WSI patches and cross-entropy to classify low-grade scores 6–7 versus high-grade scores above 7.The resulting source mapping provides the discriminative feature space for target adaptation.Source: §2 · p. 3
WSI-level prediction
- Majority voting over cropped patches produces the final Gleason score for each WSI.This aggregates patch-level predictions into the slide-level classification output.Source: §2 · p. 3
Adversarial adaptation for target domain
- Adversarial adaptation fine-tunes the target network so its feature distribution mimics the source distribution, while only the target mapping is optimized.A discriminator distinguishes source from target features, and the target mapping reverses the domain label to fool it.Source: §2 · p. 3
- The adversarial component uses a GAN loss to train the discriminator and target mapping against domain distinguishability.The discriminator receives source and target feature spaces and predicts binary domain labels.Source: §2 · p. 3
Siamese architecture at target domain
- The target-domain Siamese network pairs patches and shares weights to regularize whether patches originate from the same WSI.A one-layer perceptron classifies concatenated feature maps with cross-entropy loss, assigning 1 to same-WSI pairs and 0 otherwise.Source: §2 · p. 4
Training process
- Training proceeds in two stages: supervised source training followed by target optimization with Lt = La + Ls.The target objective combines adversarial adaptation and Siamese losses.Source: §2 · p. 4
3 Experimental Validation and Results
The method is evaluated on TCGA-based and CINJ domain shifts using binary Gleason-grade classification, with source training followed by unsupervised target adaptation. Across both shifts, adversarial adaptation with the proposed Siamese regularization improves target-domain classification over baseline approaches.
Datasets
- The experiments use TCGA institutions excluding UP as a source and UP as a target, then use all TCGA images as source and CINJ images as target.TCGA includes WSIs from 32 institutions, while CINJ images were acquired at 20× magnification and labeled with Gleason scores 6 or 8.Source: §3 · pp. 4–5
Implementation Details
- The evaluation reduces Gleason scores to low versus high grades, trains a modified fully convolutional AlexNet on source data, and then optimizes the target network with a discriminator.Source training uses scores 6–7 as low and 8–10 as high; the target network is trained while source-network parameters remain fixed.Source: §3 · pp. 5–6
Source network performance
- Both source networks outperform the prior study, although differences in dataset size and network capacity make the comparison biased.The authors use this result to indicate that the source-domain network is adequately trained for binary TCGA Gleason classification.Source: Table 3 · p. 5, §3 · p. 6
Adaptation of TCGA (w/o UP) →UP
- Unsupervised adaptation significantly improves classification for TCGA (w/o UP) →UP, with accuracy improvement over the baseline confirmed by McNemar testing (p-value 0.039).The adapted network improves classification for both low- and high-Gleason-score WSIs without additional target annotations.Source: §3 · p. 6, Table 4 · p. 6
Adaptation of TCGA →CINJ
- The proposed target-domain adaptation achieves the best performance for TCGA →CINJ when using L_t.Qualitative heatmaps show that L_t correctly classifies all patches in a high-grade CINJ WSI.Source: §3 · pp. 6–7, Table 5 · p. 6
4 Conclusion
The paper combines adversarial training with a Siamese architecture to improve target-domain Gleason classification without target labels. It reports statistically significant classification results and identifies larger datasets and broader histopathology applications as future directions.
Conclusion
- Adversarial training and a Siamese architecture improve target-network classification in an unsupervised manner.The method is intended to transfer discriminative knowledge across domains without requiring target-domain annotations.Source: §4 · p. 7
- Future work will use more extensive datasets and extend the method to a wider range of histopathology image-classification problems.Source: §4 · p. 7