Self-supervised Learning through Multi-magnification Feature Correspondence for Histopathological Image Analysis

Deep learning has shown promising results in pathological imaging analysis. However, the difficulty in acquiring large amounts of labeled data poses a significant challenge because annotating pathological images requires expert knowledge. Self-supervised learning (SSL) of unlabeled pathological images has emerged as an effective solution. The models pre-trained on the same pathological image domain as the target diagnostic task outperformed the models pre-trained on natural images. Although pathological diagnosis relies on observations across a wide range of structures, from tissue architecture to cellular details, previous SSL methods could not extract features that reflect this diagnostic process, resulting in limited performance in target diagnostic tasks. In this study, we proposed an SSL method for learning consistent feature representations across various structures in pathological images by aligning with microscopic examinations that integrate observations at multiple magnifications. Specifically, our method encourages feature representations of the same tissue region to be similar across different magnifications while maintaining the distinctive characteristics observed at each scale. Experimental evaluations of four pathological image classification tasks demonstrated that our method performs better than previous SSL methods.