- The paper introduces NearbyPatchCL, a novel self-supervised method that uses neighboring patches as positive pairs to counter imbalanced data in WSI analysis.
- It employs a decoupled contrastive loss and achieves over 87% top-1 accuracy with just 1% labeled data, highlighting its efficiency in limited annotation settings.
- The method’s success on the public P-CATCH canine cancer dataset underscores its potential for clinical integration and advances in digital pathology.
Introduction
Whole-slide image (WSI) analysis is a critical component in modern cancer diagnosis and treatment. The shift toward digitizing tissue slides has increased the need for automated and precise analysis algorithms. In this context, deep learning has taken center stage, often requiring large annotated datasets for training. Since annotations for WSI are time-consuming and expensive, self-supervised learning (SSL) methods, which do not require annotated data, have become a key interest. SSL uses a pre-training phase with unlabeled data, followed by a fine-tuned phase on a smaller labeled dataset. Although this approach is promising, the instability arises when these methods are subjected to imbalanced datasets, a common occurrence in WSI.
Self-Supervised Learning in Whole-Slide Images
Unlike traditional machine learning, SSL does not rely on extensive labeled datasets. It uses unlabeled data to learn representations before fine-tuning on a specific task with limited labeled data. Previous SSL approaches have applied random cropping to generate patches from WSIs. However, due to the diverse and skewed nature of tissue sections, this can result in imbalance problems. Furthermore, while contrastive learning has been employed to improve this process, it is not wholly resistant to issues caused by imbalanced data. To mitigate this, new methods are needed that can be effectively integrated into medical routines.
NearbyPatchCL Methodology
To address these challenges, a novel method called Nearby Patch Contrastive Learning (NearbyPatchCL) is introduced. This method uses a contrastive learning paradigm that treats adjacent patches as positive samples and employs a decoupled contrastive loss (DCL) to assist in learning robust representations. NearbyPatchCL seeks to achieve strong, stable features by viewing nearby patches as affirming, which is in contrast to existing methods that only consider different views of the same patch as positive.
What sets NearbyPatchCL apart is that it effectively benefits from leveraging neighboring patches in WSI without needing extensive annotation. This offers a way to leverage label-like information in a self-supervised setting. By enhancing performance with a fraction of labeled data, it shows great potential for real-world clinical scenarios that often struggle with limited annotations.
Evaluation and Contributions
To validate this method, the researchers curated a new dataset named P-CATCH from public WSIs focused on canine cancer. Rigorous experiments on P-CATCH demonstrated that NearbyPatchCL notably outperformed both the conventional supervised baseline and other advanced SSL methods. Specifically, it achieved top-1 classification accuracy of above 87% even when using only 1% labeled data, underscoring its efficiency.
In summary, this paper contributes a novel SSL framework that copes well with the complex challenges of imbalanced data in WSI. It suggests that NearbyPatchCL could indeed transition into clinical practice, offering a workable solution for high-quality WSI analysis with limited annotations. The research also released a new publicly available WSI dataset, fostering further development and benchmarking of patch-level multi-class classification methods in digital pathology.