[Paper Review] Self-Supervised Similarity Learning for Digital Pathology

Jacob Gildenblat, Eldad Klaiman|arXiv (Cornell University)|May 20, 2019

AI in cancer detectionComputer Science17 references33 citations

TL;DR

The paper introduces a self-supervised Siamese network that leverages spatial continuity in whole slide images to learn 128-dimension descriptors for digital pathology, outperforming ImageNet-based and other self-supervised methods on Camelyon16.

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 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. To the best of our knowledge, this is also the first published method for self-supervised learning tailored for digital pathology.

Motivation & Objective

• Motivate domain-specific feature extraction for digital pathology where labeled data are scarce.
• Leverage intrinsic spatial continuity in whole slide images to create automatically labeled pairs.
• Train a Siamese network with contrastive loss to produce compact 128-dimensional descriptors.
• Demonstrate improved image retrieval and descriptor separability over baselines on Camelyon16.

Proposed method

• Construct image pairs from tissue tiles labeled as similar if within 2 mm and non-similar if beyond 10 mm.
• Train a Siamese network with a 128-d output using contrastive loss: L_contrastive = (1-y)·||f1-f2||2 + y·max(0, m - ||f1-f2||2).
• Use a modified ResNet-50 backbone shared between branches as the encoder.
• Normalize tiles with stain normalization in LAB space before evaluation.
• Compare against ImageNet-pretrained ResNet-50 features and NPID self-supervised features.
• Evaluate via image retrieval and the global ADDR metric on Camelyon16.

Experimental results

Research questions

• RQ1 Can self-supervised similarity learning capture domain-specific histopathology features without manual labels?
• RQ2 Do descriptor-based methods using spatial continuity outperform ImageNet-based features and existing self-supervised approaches in pathology tasks?
• RQ3 How well do learned descriptors separate tumor from normal tissue in retrieval and distance-based evaluations?
• RQ4 What is the dimensionality and practical efficiency of the learned descriptors for large-scale WSI analysis?

Key findings

• Global ADDR: ImageNet ResNet-50 = 1.38, NPID = 1.28, Ours = 1.50.
• Tumor tile retrieval: ImageNet ResNet-50 = 26%, NPID = 21%, Ours = 34%.
• Our descriptors yield more clustered tumor embeddings in UMAP visualizations compared to baselines.
• Our method provides better retrieval of tumor tiles across slides, indicating domain-specific feature capture.