Source-linked AI summary
Domain-adversarial neural networks to address the appearance variability of histopathology images
Maxime W. Lafarge, Josien P. W. Pluim, Koen A. J. Eppenhof, Pim Moeskops, Mitko Veta
TL;DR
Histopathology image appearance varies across laboratories, limiting the generalization of automated analysis methods. This paper compares domain-adversarial training, color augmentation, and staining normalization for breast-cancer mitosis detection, finding that color augmentation combined with domain-adversarial training performs best.
Problem
Variability in histopathology image appearance across laboratories limits the generalization of automated image-analysis methods.
Method
The study compares domain-adversarial training with color augmentation and staining normalization for mitosis detection in breast-cancer histopathology images.
Results
On the external test set, color augmentation combined with domain-adversarial training achieved the best performance, with an F1-score of 0.62 ± 0.00 versus 0.33 ± 0.08 for the baseline.
Takeaways & Limitations
The combination’s best performance indicates that domain-adversarial training can complement color augmentation by accounting for variability beyond color.
Takeaways & Limitations
The approach assumes that each digital slide reflects a unique combination of preparation parameters.
Abstract
from arXiv · showhide
Preparing and scanning histopathology slides consists of several steps, each with a multitude of parameters. The parameters can vary between pathology labs and within the same lab over time, resulting in significant variability of the tissue appearance that hampers the generalization of automatic image analysis methods. Typically, this is addressed with ad-hoc approaches such as staining normalization that aim to reduce the appearance variability. In this paper, we propose a systematic solution based on domain-adversarial neural networks. We hypothesize that removing the domain information from the model representation leads to better generalization. We tested our hypothesis for the problem of mitosis detection in breast cancer histopathology images and made a comparative analysis with two other approaches. We show that combining color augmentation with domain-adversarial training is a better alternative than standard approaches to improve the generalization of deep learning methods.
1 Introduction
Histopathology image appearance varies because tissue preparation is inconsistent, limiting generalization—especially across pathology laboratories. The study proposes domain-adversarial training for CNN-based mitosis detection and compares it with color augmentation and staining normalization.
- Problem: Inconsistent tissue preparation causes substantial histopathology appearance variability that hampers image-analysis generalization, particularly on external-laboratory datasets.This variability arises in preparing and scanning slides and can differ between laboratories or over time.
- Existing approaches: Staining normalization standardizes colors while preserving structures but does not address variability from other sources, such as tissue fixation.The introduction presents staining normalization as a common preprocessing strategy with limited coverage of appearance variation.
- Proposed approach: The authors hypothesize that domain-adversarial training can improve CNN generalization by disregarding non-relevant appearance variability.The proposed approach imposes constraints on the learned representation to remove domain-related information.
- Experimental setting: CNN models were trained for breast-cancer mitosis detection using limited data from one pathology laboratory and evaluated on external-laboratory test data.The evaluation directly tests cross-laboratory generalization in the presence of appearance variability.
- Comparative analysis: The study compares domain-adversarial training with color augmentation and staining normalization as alternative strategies for improving generalization.The comparison is designed to assess whether domain-adversarial training provides a more systematic solution than standard preprocessing approaches.
2 Materials and Methods
The study addresses poor cross-domain generalization in breast-cancer mitosis detection by comparing spatial and color augmentation, staining normalization, domain-adversarial training, and their combinations. Models use a common CNN baseline, domain labels from source slides, and F1-score evaluation with validation-selected thresholds.
- The baseline is a 6-layer CNN that receives 63 × 63 image patches and predicts the probability of a mitotic figure at the patch center.It contains four convolutional and two fully connected layers; convolutional layers use 16 feature maps.
- Poor generalization motivates evaluating feasible alternatives because histology annotation is time-consuming and requires scarce expertise.
- All models use spatial augmentation with arbitrary rotation, mirroring, and ±20% scaling, while color augmentation targets staining-related appearance variability.Color transformations apply Ic ← ac · Ic + bc, with ac ∼ U[0.9, 1.1] and bc ∼ U[−10, +10].
- Domain-adversarial training uses slide identity as the domain label and learns representations from which the sample’s domain of origin cannot be recovered.A parallel CNN predicts domain identifiers from activations of the second and fourth mitosis-classifier layers, while the adversarial strength is controlled by α ∈ [0, 1].
- Performance is measured with F1-score from dense probability maps, using the validation-set threshold that maximizes F1-score; t-SNE qualitatively compares domain overlap.
3 Experiments and Results
Experiments compared all combinations of three approaches using repeated CNN training under a shared baseline architecture and procedure. Performance was assessed with F1-scores on internal and external test sets, alongside t-SNE visualizations of learned representations.
- Experimental design: All combinations of the three approaches were evaluated with three independently initialized convolutional neural networks using the same architecture and training procedure.Random initialization seeds were used to assess consistency across repeated experiments.
- Training procedure: Hard negative mining addressed the training set’s high class imbalance, using one shared set of hard-negative samples across all experiments.Hard negatives were sampled probabilistically from probability maps produced by an initial baseline classifier, excluding ground-truth locations.
- Domain-adversarial training: DANN training used two passes per iteration: baseline-style updates followed by domain-balanced updates for the adversarial objectives.Because the training set contained eight domains, the second-pass batches used 8 random patches from each training case.
- Domain-adversarial training: Cyclic scheduling of the adversarial parameter α was used to manage DANN instability and alternate task learning with adversarial training.The schedule was intended to prevent local maxima and avoid recovery of domain information in the main branch.
- Performance: Performance comparisons reported mean and standard-deviation F1-scores for each method and combination on internal and external test sets, with t-SNE embeddings visualizing learned features.The test sets represented the same pathology lab internally and different pathology labs externally.
4 Discussion and Conclusions
Color augmentation combined with domain-adversarial training achieved the best performance on both internal and external test sets, while domain-adversarial approaches reduced domain information in learned representations. The authors conclude that this combination is promising, but note that the study covered only one image-analysis problem and one staining-normalization method.
- Performance: Color augmentation plus domain-adversarial training achieved the best performance on both test sets.The combination obtained an internal F1-score of 0.62±0.02 and an external F1-score of 0.62 ± 0.00.
- Performance: 0.33 ± 0.08 was the external baseline F1-score, and all three investigated methods improved on it.The baseline had the worst external-test performance.
- Representation analysis: The baseline representation contained well-defined domain clusters, whereas all three approaches produced domain confusion in t-SNE embeddings.This indicates that the approaches reduced domain information in the learned representation relative to the baseline.
- Method comparison: Staining normalization improved external-lab generalization but adversely affected combinations with other methods, possibly by reducing training variability and increasing overfitting.Its combinations had worse performance than corresponding combinations without staining normalization.
- Method comparison: Color augmentation was the best individual method and produced representations more independent of the lab than the baseline, staining normalization, or DANN.This supports the observation that histopathology appearance variability is mostly manifested as staining variability.
- Conclusions and limitations: The study confirmed the relevance of domain-adversarial approaches but was limited to a single histopathology image-analysis problem and one staining-normalization method.These limitations constrain the scope of the conclusion.