Abstract
Using features extracted from networks pretrained on ImageNet is a common practice in applications of deep learning for digital pathology. However, it presents the downside of missing domain-specific image information. In digital pathology, supervised training data is expensive and difficult to collect. We propose a self-supervised method for feature extraction by similarity learning on whole slide images (WSI) that is simple to implement and allows the creation of robust and compact image descriptors. We train a siamese network, exploiting image spatial continuity and assuming spatially adjacent tiles in the image are more similar to each other than distant tiles. Our network outputs feature vectors of length 128, which allows dramatically lower memory storage and faster processing than networks pretrained on ImageNet. We apply the method on digital pathology WSIs from the Camelyon16 train set and assess and compare our method by measuring image retrieval of tumor tiles and descriptor pair distance ratio for distant/near tiles in the Camelyon16 test set. We show that our method yields better retrieval task results than existing ImageNet-based and generic self-supervised feature extraction methods.
Introduction: The Domain Gap in Digital Pathology
In digital pathology, Whole Slide Images (WSIs) are massive, often exceeding a billion pixels. Processing these images typically involves breaking them into smaller "tiles." The standard approach has been to use Convolutional Neural Networks (CNNs) pretrained on ImageNet to extract features. However, a significant "domain gap" exists: visual features of everyday objects (like animals or household items) do not necessarily translate to the subtle textures of biological tissues such as stroma, lymphocytes, or necrotic tissue. While training models directly on pathology data is ideal, obtaining supervised labels from expert pathologists is prohibitively expensive and time-consuming. This research addresses the need for domain-specific feature extraction without the burden of manual annotation.
Methodology: Learning from Spatial Continuity
The proposed method leverages the inherent spatial continuity of biological tissue. The core hypothesis is that two tiles located close to each other on a slide are likely to represent the same tissue type, while tiles far apart are likely to be different.
The Pretext Task and Architecture
The researchers defined "similar" pairs as tiles within 2mm of each other and "dissimilar" pairs as those at least 10mm apart. A Siamese Network architecture was employed, consisting of two identical ResNet-50 branches sharing weights. Unlike standard ResNet models that output 2048-dimensional vectors, this network was modified to produce a compact 128-dimensional feature vector.
Training and Loss Function
The model was trained on the Camelyon16 training set, comprising 270 WSIs of breast lymph node tissue stained with Hematoxylin and Eosin (H&E). This process generated approximately 70 million tile pairs. A contrastive loss function was used to minimize the distance between descriptors of similar pairs in the feature space while maximizing the distance between dissimilar pairs.
Results and Evaluation
The method was validated on the Camelyon16 test set (130 WSIs) using expert-annotated tumor maps to evaluate how well the learned features could distinguish tumor from normal tissue.
- Global ADDR (Average Descriptor Distance Ratio): This metric measures the separation between similar and dissimilar tiles. The proposed method achieved a score of 1.5, outperforming ImageNet-pretrained models (1.38) and generic self-supervised methods like NPID (1.28).
- Tumor Retrieval: In tasks requiring the identification of similar tumor tiles across different slides, the proposed method reached 34% accuracy, significantly higher than ImageNet (26%) and NPID (21%).
- Feature Compactness: By reducing the feature vector to 128 dimensions, the authors achieved a 16x reduction in storage requirements and faster downstream processing without compromising performance.
Discussion and Conclusion
The "geometry as a label" approach effectively utilizes the physical structure of the slide as a free labeling system. However, the spatial continuity assumption is not without flaws; label noise can occur at the boundaries between tumor and healthy tissue where nearby tiles may differ significantly. Future work may involve refining proximity thresholds and exploring more sophisticated sampling strategies to mitigate this noise. Despite these challenges, the results demonstrate that domain-specific self-supervised learning provides a robust and efficient alternative to general pretraining for digital pathology applications.
Summary updated