Abstract
This work presents a novel self-supervised pre-training method to learn efficient representations without labels on histopathology medical images utilizing magnification factors. Other state-of-the-art works mainly focus on fully supervised learning approaches that rely heavily on human annotations. However, the scarcity of labeled and unlabeled data is a long-standing challenge in histopathology. Currently, representation learning without labels remains unexplored in the histopathology domain. The proposed method, Magnification Prior Contrastive Similarity (MPCS), enables self-supervised learning of representations without labels on small-scale breast cancer dataset BreakHis by exploiting magnification factor, inductive transfer, and reducing human prior. The proposed method matches fully supervised learning state-of-the-art performance in malignancy classification when only 20% of labels are used in fine-tuning and outperform previous works in fully supervised learning settings for three public breast cancer datasets, including BreakHis. Further, It provides initial support for a hypothesis that reducing human-prior leads to efficient representation learning in self-supervision, which will need further investigation.
Introduction
In digital pathology, Whole Slide Images (WSIs) are massive, high-resolution files that present a significant labeling challenge. While deep learning has shown promise in cancer diagnosis, most models are "data-hungry," relying on supervised learning that is often bottlenecked by the scarcity of expert-annotated medical data. Existing self-supervised learning (SSL) methods typically rely on generic human-inducted priors—such as artificial rotations or color shifts—which may not capture the unique hierarchical structure of medical tissue. This paper introduces Magnification Prior Contrastive Similarity (MPCS), a method that leverages the inherent properties of histopathological data—specifically magnification—to learn robust representations with minimal human intervention.
Methodology
The MPCS framework operates through a three-stage pipeline designed to transition from cross-domain to domain-specific representations.
- Inductive Transfer: The model is initialized with weights from ImageNet to establish a baseline understanding of general visual features.
- Self-Supervised Pre-training (MPCS): This core stage uses different magnification levels (e.g., 40x, 100x, 200x, 400x) of the same tissue sample as "positive pairs" in a contrastive learning framework. By maximizing the similarity between different views of the same sample, the model learns features that are invariant to zoom levels.
- Fine-tuning: The pre-trained encoder is then fine-tuned on a small subset of labeled data for downstream tasks like malignancy classification.
A critical aspect of this research is the exploration of "Human Priors" versus "Degrees of Freedom" (DoF) in sampling. The authors tested three strategies: Fixed Pair (0 DoF, high human prior), Ordered Pair (1 DoF, moderate human prior), and Random Pair (2 DoF, low human prior). Interestingly, the "Random Pair" strategy, which allows the model the most freedom in selecting magnification combinations, generally yielded the most efficient representations.
Experimental Results
The study evaluated MPCS on three public datasets: BreakHis, BACH (Breast Cancer Histology), and the Breast Cancer Cell Dataset.
Performance and Label Efficiency
MPCS demonstrated remarkable label efficiency. On the BreakHis dataset, the model achieved malignancy classification accuracy comparable to state-of-the-art supervised models while utilizing only 20% of the labels. In fully supervised settings (100% labels), MPCS outperformed previous benchmarks across all three datasets, achieving approximately 91.85% image-wise accuracy on the BACH dataset.
Qualitative Analysis
To understand what the model learned, the authors employed Class Activation Maps (CAM). As shown in the comparison below, models pre-trained with MPCS (specifically the Ordered Pair strategy) focus more accurately on relevant cellular structures, such as nuclei clusters, whereas models pre-trained solely on ImageNet often focus on irrelevant background noise or stroma.
Discussion and Conclusion
The findings suggest that magnification acts as a powerful natural supervision signal in pathology. Furthermore, the success of the "Random Pair" strategy supports the hypothesis that reducing human-inducted priors allows the model to discover more robust, data-driven features.
While the current method utilizes discrete magnification levels, future research could explore learning on a continuous magnification scale. Additionally, testing MPCS on larger, multi-center clinical datasets will be essential for validating its utility in real-world diagnostic workflows. This work provides a significant step toward efficient, self-supervised representation learning in the medical domain, where labeled data remains a precious resource.
Summary updated