[Paper Review] Magnification Prior: A Self-Supervised Method for Learning Representations on Breast Cancer Histopathological Images

Prakash Chandra Chhipa, Richa Upadhyay+4 more|arXiv (Cornell University)|Mar 15, 2022

AI in cancer detection4 citations

TL;DR

This paper proposes Magnification Prior Contrastive Similarity (MPCS), a self-supervised representation learning method for breast cancer histopathology images that leverages magnification factors as a data-derived supervisory signal to create positive pairs without human-annotated labels. MPCS achieves state-of-the-art performance on malignancy classification with only 20% labeled data, outperforming fully supervised methods and demonstrating that reducing human-induced priors enhances representation learning efficiency.

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-theart 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 for 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. It formulates a hypothesis and provides empirical evidence to support that reducing human-prior leads to efficient representation learning in self-supervision. The implementation of this work is available online on GitHub - https://github.com/prakashchhipa/Magnification-Prior-Self-Supervised-Method

Motivation & Objective

• To address the scarcity of labeled data in histopathology image analysis by enabling self-supervised pre-training without human annotations.
• To explore whether data-derived priors—specifically magnification factors—can serve as effective supervision signals for contrastive representation learning.
• To investigate the hypothesis that reducing human-induced priors in self-supervised learning improves representation quality and generalization.
• To evaluate the effectiveness of MPCS on small-scale, public breast cancer datasets including BreakHis, BACH, and Breast Cancer Cell Dataset.
• To demonstrate label efficiency by matching or exceeding fully supervised SOTA performance with only 20% of labels.

Proposed method

• MPCS uses magnification factors (40X, 100X, 200X, 400X) as a data prior to generate positive pairs for contrastive learning, eliminating reliance on human-designed augmentations.
• The method employs three variants: MPCS-Random Pair (random view sampling), MPCS-Ordered Pair (ordered sampling), and MPCS-Fixed Pair (fixed 200X–400X pairs), varying degrees of human prior induction.
• Contrastive learning is applied using a Siamese network architecture with feature encoders pre-trained on ImageNet (EfficientNet and dilated ResNet-50).
• Self-supervised pre-training is performed on unlabeled BreakHis images before fine-tuning on downstream malignancy classification tasks.
• The model is evaluated via linear evaluation and fine-tuning on three public datasets: BreakHis, BACH, and Breast Cancer Cell Dataset.
• Class activation mapping (CAM) is used to qualitatively validate that MPCS models learn to attend to biologically relevant regions in histopathology images.

Figure 1 : The proposed approach comprises three steps: (1) Parameters initialization with supervised ImageNet weights. (2) Self-supervised pre-training on unlabeled BreakHis histopathology images using proposed method Magnification Prior Contrastive Similarity to provide positive pairs by exploitin

Experimental results

Research questions

• RQ1 Can magnification factors serve as a reliable data prior to enable self-supervised contrastive learning in histopathology without human-annotated labels?
• RQ2 Does reducing human-induced prior in pair sampling improve the quality of learned representations in self-supervised representation learning?
• RQ3 How does MPCS perform in terms of label efficiency compared to fully supervised state-of-the-art methods?
• RQ4 Can MPCS learn discriminative features across benign and malignant classes without access to ground-truth labels during pre-training?
• RQ5 Does cross-magnification evaluation confirm that MPCS representations are magnification-invariant and robust?

Key findings

• MPCS outperforms fully supervised SOTA methods on three public datasets, including BreakHis, when fine-tuned with only 20% of the labels.
• On the Breast Cancer Cell Dataset, MPCS achieves 98.18% accuracy in fine-tuning and 96.36% in linear evaluation, significantly outperforming prior methods.
• With only 20% labeled data, MPCS fine-tuned models achieve a 2.52% absolute improvement over ImageNet pre-trained models (p < 0.01), matching SOTA performance trained on 100% labels.
• The MPCS-Ordered Pair variant, which induces weaker human prior, outperforms both the fixed-pair (stronger human prior) and random-pair (no human prior) variants under low-label regimes.
• t-SNE visualization confirms that MPCS pre-trained features are highly discriminative and separable across benign and malignant classes across all magnification levels.
• CAM analysis shows that MPCS-fine-tuned models activate regions of interest (e.g., nuclei) more effectively than ImageNet-pretrained models on BreakHis and BACH datasets.

Figure 2 : Magnification Prior Contrastive Similarity method explained