Source-linked AI summary
Context-aware stacked convolutional neural networks for classification of breast carcinomas in whole-slide histopathology images
Babak Ehteshami Bejnordi, Guido Zuidhof, Maschenka Balkenhol, Meyke Hermsen, Peter Bult, Bram van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen van der Laak
TL;DR
Breast WSI classification must capture both cellular detail and broad tissue architecture to distinguish normal/benign tissue, DCIS, and IDC. The paper introduces a context-aware stacked CNN that combines high-resolution patch features with larger contextual inputs, achieving strong binary and three-class performance on 221 WSIs. Its main scope boundary is the remaining difficulty of separating visually similar benign lesions from DCIS and DCIS from IDC.
Problem
Distinguishing benign lesions, DCIS, and IDC requires cellular detail and broader tissue architecture, but computational constraints limit direct training on large high-resolution patches.
Method
The method trains a CNN on small high-resolution patches, stacks a second CNN on its fixed features, and uses larger patches to incorporate contextual tissue structure.
Results
The system achieved an AUC of 0.962 for binary normal/benign-versus-cancer classification and 0.812 accuracy for three-class WSI classification.
Takeaways & Limitations
The results support context-aware stacked CNNs as a potential approach for breast pathology diagnostics, including screening out approximately 50% of obviously normal/benign slides without missing cancerous slides on the dataset.
Takeaways & Limitations
Errors remained common for benign-versus-DCIS and DCIS-versus-IDC discrimination, with performance limited by similar appearances and the need for larger receptive fields.
Abstract
from arXiv · showhide
Automated classification of histopathological whole-slide images (WSI) of breast tissue requires analysis at very high resolutions with a large contextual area. In this paper, we present context-aware stacked convolutional neural networks (CNN) for classification of breast WSIs into normal/benign, ductal carcinoma in situ (DCIS), and invasive ductal carcinoma (IDC). We first train a CNN using high pixel resolution patches to capture cellular level information. The feature responses generated by this model are then fed as input to a second CNN, stacked on top of the first. Training of this stacked architecture with large input patches enables learning of fine-grained (cellular) details and global interdependence of tissue structures. Our system is trained and evaluated on a dataset containing 221 WSIs of H&E stained breast tissue specimens. The system achieves an AUC of 0.962 for the binary classification of non-malignant and malignant slides and obtains a three class accuracy of 81.3% for classification of WSIs into normal/benign, DCIS, and IDC, demonstrating its potentials for routine diagnostics.
1 Introduction
Breast WSI classification must distinguish normal/benign tissue, DCIS, and IDC despite difficult visual overlap and the scale of microscopic data. The study therefore develops a context-aware stacked CNN using high-resolution cellular information and broader tissue context.
- Accurate discrimination of DCIS and IDC from benign breast disease is important for determining treatment plans.
- Microscopic assessment is tedious, time-consuming, subjective, and increasingly burdensome as screening expands the volume of tissue data.
- Existing breast histopathology algorithms commonly analyze the morphology and spatial arrangement of epithelial structures such as nuclei and ducts.
- CNNs learn hierarchical feature representations directly from data and have motivated deep-learning approaches to breast cancer image classification.
- The study targets three-class WSI classification because benign lesions vary widely and DCIS can visually resemble IDC at the cellular level.
- The proposed system stacks a high-resolution CNN with a second CNN to combine cellular features and global structural interdependence across 221 breast WSIs.
2 Methods
The method sequentially analyzes small high-resolution patches and larger contextual patches with stacked CNNs, then aggregates lesion-level spatial features for WSI classification. Fixed lower-stage weights make larger effective patch sizes computationally feasible.
- 2.1 Overview of the system: DCIS and IDC may look similar at the cellular level, so classification requires larger patches that capture global tissue architecture.
- 2.1 Overview of the system: Computational constraints make it infeasible to train a deep CNN directly on large, high-resolution patches containing sufficient context.
- 2.2 Deep CNN for classification of small high-resolution patches: WRN-4-2 classifies 224 × 224 high-resolution patches using a wide residual architecture with N=4 and K=2.
- 2 Methods: The first CNN’s learned convolutional features are transferred to a stacked network for contextual prediction.
- 2.3 Context-aware Stacked CNN (CAS-CNN): CAS-CNN stacks a second fully convolutional CNN above WRN-4-2, fixes the first network’s parameters, and trains with 512 × 512, 768 × 768, and 1024 × 1024 patches.
- 2 Methods: Dense prediction is produced by sliding the stacked network over each WSI with a stride of 224.
- 2 Methods: Prediction maps are converted into lesion labels, connected components, global area features, convex-hull features, and spatial-distribution descriptors.
- 2 Methods: The resulting 57 features are classified with two random forests, using 512 decision trees for three-class and binary WSI classification.
3 Experiments
The experiments evaluate context-aware CNNs for breast WSI classification using slide-level splits, patch sampling, augmentation, and hard-negative mining. The system achieves strong binary and three-class test performance, while larger training patches improve validation accuracy but increase computational cost.
- Dataset and evaluation: 221 H&E-stained WSIs from 122 patients were split into non-overlapping training, validation, and test sets at both slide and patient levels.The dataset included 100 normal/benign, 69 DCIS, and 55 IDC WSIs; the test set contained 64 WSIs.
- Dataset and evaluation: The evaluation measured accuracy and Cohen’s kappa for three-class WSI classification and AUC, accuracy, and kappa for binary normal/benign-versus-cancer classification.Cancer combined DCIS and IDC for the binary task.
- Intermediate validation: 0.924 patch-level binary accuracy exceeded 0.799 three-class accuracy for WRN-4-2 on validation data, consistent with limited context from 224 × 224 patches.The reported explanation is that small patches lack enough context to discriminate the three classes accurately.
- Intermediate validation: 0.872, 0.905, and 0.914 three-class validation accuracies were obtained with CAS-CNN trained on 512 × 512, 768 × 768, and 1024 × 1024 patches, respectively.The 768 × 768 model was selected for dense prediction because the 1024 × 1024 network had heavy computational costs.
- Whole-slide results: 0.962 AUC was achieved for binary test-set WSI classification, with 0.891 accuracy and 0.781 kappa.The selected 768 × 768 CAS-CNN produced dense prediction maps before random-forest WSI classification.
- Whole-slide results: 0.813 accuracy and 0.700 kappa were achieved for three-class test-set classification of normal/benign, DCIS, and IDC WSIs.The result was reported for the overall three-class system.
4 Discussion and Conclusion
The paper presents CAS-CNN as a two-stage architecture combining cellular-level features with global tissue-structure context for breast WSI classification. It reports strong binary discrimination and promising three-class performance, while identifying errors, computational costs, and stain variation as important boundaries.
- Architecture: CAS-CNN combines cellular-level features from small high-resolution patches with global interdependence of tissue structures from larger inputs.The architecture stacks a fully convolutional network on top of the first CNN.
- Performance: CAS-CNN trained on large input patches outperformed a wide ResNet using 224 × 224 patches and improved consistently with larger input patches.The authors connect this result to incorporating more context into classification.
- Performance: An AUC of 0.962 was achieved for binary classification of normal/benign versus cancerous slides.The dataset contained 10 benign categories, and the authors report potential to sieve out approximately 50% of obviously normal/benign slides without missing cancerous slides.
- Performance: An accuracy of 0.812 and a Kappa value of 0.700 were achieved for three-class WSI classification.The classes were normal/benign, DCIS, and IDC.
- Limitations: Errors were common for normal/benign versus DCIS and DCIS versus IDC discrimination, with larger receptive fields proposed as one possible remedy.The paper also notes that increasing patch size improved performance but increased training and inference time.
- Limitations: Stain standardization is identified as a likely way to improve network generalization across whole-slide images.The paper notes that specialized stain standardization techniques have improved computer-aided diagnostic performance by reducing stain variation.
- Future work: The system is presented as a first step toward finding prognostic and predictive biomarkers within malignant lesions.Biomarker discovery is identified as a direction for future work beyond lesion-category discrimination.