Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unsupervised Latent Stain Adaptation for Computational Pathology

Daniel Reisenbüchler, Lucas Luttner, Nadine S. Schaadt, Friedrich Feuerhake, Dorit Merhof

TL;DR

Unsupervised Latent Stain Adaptation (ULSA) addresses the problem of staining-induced domain shifts in computational pathology by jointly leveraging annotated source-stain data, artificially translated target-stain data via a cGAN, and unlabeled target-stain images. The method combines (i) stain-translation augmentation to enrich labeled data, (ii) an unsupervised stain adaptation objective that blends supervised and unsupervised losses, and (iii) stain-invariant feature consistency learning that aligns hierarchical latent representations across stains. The training objective minimizes a joint loss $ \min_{\\theta} \\mathcal{L}(\\theta|x^{L},y,x^{U}) = \\mathbb{E}_{x^{L}\\sim \\hat{p}_{S\\cup T}} [\\mathcal{L}_{S}(\\theta|x^{L},y)] + \\lambda \\mathbb{E}_{x^{U}\\sim p_{S\\cup T}} [\\mathcal{L}_{U}(\\theta|x^{U})]$, enabling efficient stain adaptation for segmentation and classification tasks. Empirically, ULSA achieves state-of-the-art performance on kidney tissue segmentation and breast cancer classification across diverse stains, performs well with only 10% of labels, and demonstrates robustness as a task-agnostic, patch-level framework suitable for WSIs and scalable to multiple staining variations.

Abstract

In computational pathology, deep learning (DL) models for tasks such as segmentation or tissue classification are known to suffer from domain shifts due to different staining techniques. Stain adaptation aims to reduce the generalization error between different stains by training a model on source stains that generalizes to target stains. Despite the abundance of target stain data, a key challenge is the lack of annotations. To address this, we propose a joint training between artificially labeled and unlabeled data including all available stained images called Unsupervised Latent Stain Adaptation (ULSA). Our method uses stain translation to enrich labeled source images with synthetic target images in order to increase the supervised signals. Moreover, we leverage unlabeled target stain images using stain-invariant feature consistency learning. With ULSA we present a semi-supervised strategy for efficient stain adaptation without access to annotated target stain data. Remarkably, ULSA is task agnostic in patch-level analysis for whole slide images (WSIs). Through extensive evaluation on external datasets, we demonstrate that ULSA achieves state-of-the-art (SOTA) performance in kidney tissue segmentation and breast cancer classification across a spectrum of staining variations. Our findings suggest that ULSA is an important framework for stain adaptation in computational pathology.

Unsupervised Latent Stain Adaptation for Computational Pathology

TL;DR

Unsupervised Latent Stain Adaptation (ULSA) addresses the problem of staining-induced domain shifts in computational pathology by jointly leveraging annotated source-stain data, artificially translated target-stain data via a cGAN, and unlabeled target-stain images. The method combines (i) stain-translation augmentation to enrich labeled data, (ii) an unsupervised stain adaptation objective that blends supervised and unsupervised losses, and (iii) stain-invariant feature consistency learning that aligns hierarchical latent representations across stains. The training objective minimizes a joint loss , enabling efficient stain adaptation for segmentation and classification tasks. Empirically, ULSA achieves state-of-the-art performance on kidney tissue segmentation and breast cancer classification across diverse stains, performs well with only 10% of labels, and demonstrates robustness as a task-agnostic, patch-level framework suitable for WSIs and scalable to multiple staining variations.

Abstract

In computational pathology, deep learning (DL) models for tasks such as segmentation or tissue classification are known to suffer from domain shifts due to different staining techniques. Stain adaptation aims to reduce the generalization error between different stains by training a model on source stains that generalizes to target stains. Despite the abundance of target stain data, a key challenge is the lack of annotations. To address this, we propose a joint training between artificially labeled and unlabeled data including all available stained images called Unsupervised Latent Stain Adaptation (ULSA). Our method uses stain translation to enrich labeled source images with synthetic target images in order to increase the supervised signals. Moreover, we leverage unlabeled target stain images using stain-invariant feature consistency learning. With ULSA we present a semi-supervised strategy for efficient stain adaptation without access to annotated target stain data. Remarkably, ULSA is task agnostic in patch-level analysis for whole slide images (WSIs). Through extensive evaluation on external datasets, we demonstrate that ULSA achieves state-of-the-art (SOTA) performance in kidney tissue segmentation and breast cancer classification across a spectrum of staining variations. Our findings suggest that ULSA is an important framework for stain adaptation in computational pathology.
Paper Structure (19 sections, 3 equations, 3 figures, 1 table)

This paper contains 19 sections, 3 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. (i) cGAN stain augmentation.
  4. (ii) Unsupervised stain adaptation.
  5. (iii) Stain-invariant feature consistency learning.
  6. Experiments
  7. Datasets and Comparable Methods
  8. Kidney segmentation datasets.
  9. Breast cancer classification datasets.
  10. Comparable methods.
  11. Implementation
  12. cGAN augmentation.
  13. Segmentation and classification.
  14. Results and Discussion
  15. Measuring stain adaptation.
  16. ...and 4 more sections

Figures (3)

  • Figure 1: (a). Problem statement of unsupervised stain adaptation. (b). Stain-invariant feature consistency learning. (c). Artificial images generated by cGAN.
  • Figure 2: ULSA model. Labeled source stains are translated into synthetic target stain data to obtain supervision for image-wise stain-invariance. We extract features for a real and stain translated noised image of the target stain data, where we maximize cosine similarity to achieve unsupervised feature-wise stain-invariance in latent space.
  • Figure 3: (a). Performance for different fractions of labeled data. (b). Ablation study.