DARC: Deep adaptive regularized clustering for histopathological image classification.

Area of Science:

• Computational pathology
• Machine learning in medicine
• Digital pathology

Background:

• Deep learning excels in histopathological image classification but requires extensive manual annotations.
• Annotation is time-consuming, costly, and limits the use of unlabeled pathological images.

Purpose of the Study:

• To introduce a self-supervised framework, Deep Adaptive Regularized Clustering (DARC), to pre-train neural networks for histopathological image classification.
• To reduce the dependency on large, annotated datasets.

Main Methods:

• DARC iteratively clusters learned representations and uses cluster assignments as pseudo-labels for network training.
• An objective function combines network and clustering losses with an adaptive regularization function for discriminative representations.

Main Results:

• Pre-training with DARC significantly boosts accuracy compared to training from scratch.
• Using DARC pre-trained weights with only 10% labeled data achieves accuracy comparable to 100% labeled data training.
• DARC pre-trained networks exhibit faster convergence on downstream tasks.

Conclusions:

• DARC effectively mitigates the need for extensive data annotation in histopathological image classification.
• The learned representations are generalizable and discriminative, as shown by t-SNE visualization.
• DARC offers a promising approach for leveraging unlabeled data in computational pathology.