Source-linked AI summary
Resolving challenges in deep learning-based analyses of histopathological images using explanation methods
Miriam Hägele, Philipp Seegerer, Sebastian Lapuschkin, Michael Bockmayr, Wojciech Samek, Frederick Klauschen, Klaus-Robert Müller, Alexander Binder
TL;DR
Deep-learning pathology analyses need explanations that reveal whether models rely on medically meaningful features rather than latent dataset biases. The paper applies pixel-wise explanation heatmaps to tumour classification and evaluates them at cell level across pathology datasets. These heatmaps support quantitative inspection and bias analysis, with a reported 5% AUC improvement when reducing a labeling bias, while interpretation remains limited for surrounding tissue and spatial context.
Problem
Digital pathology needs explanation beyond predictive performance because datasets contain latent biases that can affect model generalization and real-world use.
Method
The study trains binary tumour-versus-healthy classifiers and uses Layer-wise Relevance Propagation to generate pixel-level heatmaps evaluated quantitatively at cell level.
Results
5% AUC improvement was observed when reducing a labeling bias, while heatmaps also enabled cell-level evaluation and helped reveal latent dataset biases.
Takeaways & Limitations
High-resolution explanation heatmaps can support model inspection and the detection and subsequent removal of latent biases in digital pathology analyses.
Takeaways & Limitations
Visual explanations do not unambiguously depict cell composition, adjacent tissue, or spatial relations, so ROC analysis measures relevance on cells rather than surrounding tissue.
Abstract
from arXiv · showhide
Deep learning has recently gained popularity in digital pathology due to its high prediction quality. However, the medical domain requires explanation and insight for a better understanding beyond standard quantitative performance evaluation. Recently, explanation methods have emerged, which are so far still rarely used in medicine. This work shows their application to generate heatmaps that allow to resolve common challenges encountered in deep learning-based digital histopathology analyses. These challenges comprise biases typically inherent to histopathology data. We study binary classification tasks of tumor tissue discrimination in publicly available haematoxylin and eosin slides of various tumor entities and investigate three types of biases: (1) biases which affect the entire dataset, (2) biases which are by chance correlated with class labels and (3) sampling biases. While standard analyses focus on patch-level evaluation, we advocate pixel-wise heatmaps, which offer a more precise and versatile diagnostic instrument and furthermore help to reveal biases in the data. This insight is shown to not only detect but also to be helpful to remove the effects of common hidden biases, which improves generalization within and across datasets. For example, we could see a trend of improved area under the receiver operating characteristic curve by 5% when reducing a labeling bias. Explanation techniques are thus demonstrated to be a helpful and highly relevant tool for the development and the deployment phases within the life cycle of real-world applications in digital pathology.
Introduction
Deep learning can support increasingly demanding histopathological image analysis, but predictive performance alone does not provide the explanations and insight needed in medicine. Digital pathology datasets also contain label, staining, class-balance, data-availability, and latent-bias challenges that require careful collaboration between medical and machine-learning experts.
- Deep learning offers computational support for histopathological diagnosis as diagnostic throughput and quantitative tissue assessment needs increase.
- Medical applications require explanations and insight beyond standard predictive-performance evaluation.
- Digital pathology faces noisy labels, stain variance, imbalanced classes, and limited labelled data.
- These challenges make datasets prone to latent biases that models may encode as artefactual structures.
- Addressing such biases requires close collaboration because medical experts provide labels while machine-learning experts may have limited medical-domain knowledge.
Related work
Prior work established visual explanation methods in computer vision and began applying them to medical and digital pathology images. This paper situates its approach within efforts to interpret classifier decisions and address pathology-specific challenges.
- Visual explanation methods have been successfully applied to natural images and show promise for medical imaging.
- Only a limited number of explanation-method studies had been applied in digital pathology.
- Earlier digital-pathology work explored explanations for support-vector-machine decisions using Bag-of-Visual-Words features.
- The study follows earlier work by examining explanation methods specifically for challenges in digital pathology.
Method
The study applies convolutional neural networks to tumour-versus-healthy tissue classification across three TCGA projects and explains decisions using pixel-level Layer-wise Relevance Propagation heatmaps. It evaluates these heatmaps quantitatively at cell level and aggregates them for larger tiles.
- Datasets: The experiments use H&E images from TCGA melanoma, breast carcinoma, and lung adenocarcinoma projects.
- Convolutional neural network training: Separate GoogLeNet models are fine-tuned for each tumour entity in binary classification tasks.
- Explaining classifier decisions: Layer-wise Relevance Propagation assigns pixel-level relevance scores by distributing the classifier output backward through the network.
- Explaining classifier decisions: Heatmaps use negative relevance for evidence against a class and positive relevance for evidence supporting it, with normalized values in [-1,1].
- Explaining classifier decisions: The method combines the ε-rule in the fully connected layer with the αβ-rule in succeeding convolutional layers.
- Explaining classifier decisions: Tile-level heatmap normalization preserves patch relevance peaks and their overall relationships when aggregating fixed-size neural-network heatmaps.
- Quantitative evaluation of explanation heatmaps: Cell-level ROC analysis measures cancer-cell detection sensitivity against false-positive rate across positive-relevance thresholds.
Experimental results
Across five experiments, pixel-level explanation heatmaps enabled cell-level evaluation and exposed dataset, class-correlated, and sampling biases in histopathological classification. The heatmaps also supported bias reduction, including a 5% AUC improvement after counteracting dataset bias.
- Verifying learned features: Heatmaps extended evaluation from patch-level predictions to cell-level verification of learned features across breast, skin, and lung tumour entities.Weighted F1-scores on held-out patients ranged from 0.87 to 0.95, and cell-level ROC/AUC analysis tested whether positive relevance overlapped expert-provided cancer-cell labels.
- Verifying learned features: Cell-level ROC analysis showed that computed heatmaps surpassed zero, one, and random-heatmap baselines for all three tumour entities.Zero-only and one-only heatmaps had AUC=0.5, while randomly generated heatmaps had AUC=0.5±0.0002 over 100 runs.
- Uncovering biases: High-resolution heatmaps exposed biases that common accuracy measures can miss, including dataset-wide, class-correlated, and sampling biases.The analyses examined whether relevance concentrated on unintended image regions or reflected missing structures in the training data.
- Dataset bias: 5% AUC improvement followed spatial-translation augmentation that counteracted centre-focused dataset bias.Without augmentation, absolute relevance concentrated near patch centres; with augmentation, relevance spread more evenly across the patch.
- Sampling bias: Sampling bias produced positive relevance on necrotic regions in half of the evaluated regions and greater fluctuations in mean relevance than the unbiased classifier.The comparison used five tiles containing one to five differently sized necrotic areas.
Discussion
Pixel-wise explanation heatmaps provide fine-grained views of classifier decisions that help evaluate relevant cells, reveal hidden biases, and compare explanatory methods. Their practical value is demonstrated across bias detection, model assessment, and deployment-oriented analysis, while interpretation remains limited when surrounding tissue and spatial relations matter.
- Benefits of visual explanations: Pixel-wise heatmaps move evaluation from patch level to cell-relevant resolution, while retaining tissue context as an important factor in classification.The paper uses cell-level relevance to assess whether models consider occurrences of relevant cell types, while acknowledging that adjacent tissue also contributes to decisions.
- Bias detection: High-resolution heatmaps can expose dataset, class-correlated, and sampling biases that standard accuracy measures may not reveal on individual samples.Examples include centre-cell label bias, small artificial artefacts, and missing necrotic-tissue examples.
- Practical implications: Cell-level evaluation revealed tumour-infiltrating lymphocytes as a source of relevance and raised AUC to 94% after excluding two affected tiles.The same heatmaps also reflected precision–recall trade-offs, allowing sampling strategies to emphasize the application’s target objective.
- Granularity and methods: High-resolution heatmaps offer more precise and flexible analysis than patch-level maps because small structures can be localized and evaluated at multiple scales.The comparison includes probability maps, Grad-CAM, and LRP, with probability maps having the lowest native spatial resolution.
- Limitations: Visual explanations are limited because relevance maps do not unambiguously depict cell composition, surrounding tissue, and spatial relations in an intuitive human-readable form.Accordingly, the study measures relevance on cells rather than surrounding tissue for ROC analysis.
Conclusion
The study concludes that pixel-wise visual explanations can both compare learned features with expert labels and uncover latent dataset biases. Across three tumour entities, LRP heatmaps supported cell-level evaluation and exposed information about training procedures relevant to medical applications.
- Conclusion: Pixel-wise visual explanations support quantitative and qualitative comparison of end-to-end learned features against domain-expert labels and can uncover latent dataset biases.The conclusion presents these as the two principal benefits of the approach.
- Conclusion: LRP heatmaps produced reasonable visualisations of classifier decisions across three tumour entities, with high overlap with domain-expert labels.Overlap was evaluated using ROC curves at cell level under the assumption that an ideal model considers all tumour cells.
- Conclusion: The heatmaps also reflected training-procedure knowledge such as precision–recall behaviour, supporting adjustment toward recall or precision according to application goals.The conclusion links this flexibility to medical application requirements.
APPENDIX A CONFUSION MATRICES
The appendix presents confusion matrices for binary cancer-versus-healthy tissue classifiers across three tumour entities. Interpretation should account for differing label distributions among the datasets.
- APPENDIX A CONFUSION MATRICES: Confusion matrices summarize cancerous-versus-healthy tissue classification for BRCA, LUAD, and SKCM.The entities are invasive breast cancer, lung adenocarcinoma, and cutaneous malignant melanoma.
- APPENDIX A CONFUSION MATRICES: Comparisons across the matrices should consider that the three datasets have different label distributions.The figure caption explicitly cautions against reading the results without this context.
APPENDIX B DETAILED INFORMATION ON AVAILABLE ANNOTATIONS FOR EVALUATION ON CELL LEVEL
The appendix provides exemplary annotated sample tiles from the BRCA, LUAD, and SKCM projects for cell-level evaluation. The figures identify the example tiles and, for SKCM, mark the subset excluding tiles containing tumour-infiltrating lymphocytes.
- BRCA annotations: BRCA sample tiles are shown with extensive single-cell annotations, and a blue box identifies the example used in the paper.The annotations support cell-level evaluation.
- LUAD annotations: LUAD sample tiles are shown with extensive single-cell annotations, with a blue box marking the example presented in the paper.The figure documents the annotated material used for evaluation.
- SKCM annotations: SKCM sample tiles are shown with extensive single-cell annotations, and black marks identify the subset excluding two tiles containing tumour-infiltrating lymphocytes.A blue box marks the example shown in the paper.
- Annotation summary: Table 1 summarizes the available annotations used for cell-level evaluation.The appendix section identifies the table as a summary of available evaluation annotations.
APPENDIX C DATA SAMPLING STRATEGIES
Table 2 reports model performances across different sampling ratios.
- APPENDIX C DATA SAMPLING STRATEGIES: Model performance is evaluated under different sampling ratios.The passage does not report the individual ratios or performance values.
- APPENDIX C DATA SAMPLING STRATEGIES: Sampling ratio is the comparison variable in the reported model evaluations.
- APPENDIX C DATA SAMPLING STRATEGIES: The table concerns performances of models trained with varying sampling configurations.
I. Dataset bias
The dataset-bias analysis uses relevance heatmaps and independent-test evaluation to examine how center bias affects classifier behavior. Counteracting the bias through random translations improves the ROC curve on held-out tiles.
- I. Dataset bias: The biased classifier concentrates absolute relevance predominantly in the heatmap center, unlike the unbiased classifier.
- I. Dataset bias: Mean absolute relevance is evaluated across centered relative areas on an independent test dataset unaffected by the training bias.
- I. Dataset bias: The ROC curve improves on seven held-out 1000x1000px tiles when random translations counteract the bias.
II. Sampling bias
The sampling-bias analysis compares relevance assigned to necrosis regions and visualizes classifier heatmaps on exemplary necrosis samples. It also presents average cancer heatmaps by predicted class.
- II. Sampling bias: Average relevance per annotated necrosis region is compared between classifiers on exemplary necrosis tiles.
- II. Sampling bias: Exemplary H&E necrosis samples are shown with heatmaps from classifiers trained with and without necrosis samples.
- II. Sampling bias: Average cancer heatmaps are presented separately for images classified as cancer and non-cancer.
APPENDIX E ROC CURVE FOR GRADCAM
The appendix presents GradCam heatmaps through ROC curves evaluated across all three studied tumor entities.
- APPENDIX E ROC CURVE FOR GRADCAM: ROC curves are generated from GradCam heatmaps across all three studied tumor entities.
- APPENDIX E ROC CURVE FOR GRADCAM: The figure compares receiver operating characteristic curves across the studied tumor entities.
- APPENDIX E ROC CURVE FOR GRADCAM: All three studied tumor entities are included in the GradCam-based ROC analysis.