Decomposition based curriculum-style self-training for source-free universal domain adaptation in computational...

Computational pathology models serve as crucial tools for clinical tasks such as tissue typing, alleviating the burden of manual screening of whole slide images. The stringent ethical regulations on source data, along with agnostic covariate and label co-shifts, substantially hinder the cross-institute deployment. In this dilemma, recently emerging source-free universal domain adaptation (SF-UniDA) aims to achieve knowledge transfer using only unlabeled target data, as well as common-private categories separation. To assign pseudo-labels for self-training, existing methods employ uncertainty thresholds, auxiliary networks, or self-supervised tasks. However, the internal variations within private data are overlooked by treating them as a unified whole. Besides, the interference from noisy samples remains underserved. In this paper, we design a Decomposition based Curriculum-style Self-Training (DCST) framework, to rethink SF-UniDA from data scrutiny and learning paradigm views. Specifically, an adaptive division strategy is proposed, in which a two-component Gaussian Mixture Model is empirically estimated based on the orthogonal projections of target features to derive easy/hard data boundaries. Subsequently, the easy curriculum identifies proper target clustering based on the proposed semantic and instance-level metrics, while maintaining inter-class variability within private data via prototype-aware regularization. The following hard curriculum further introduces prototype-aware alignment and easy data replay, to balance learning plasticity and memory stability. We evaluate our DCST against the state-of-the-art methods on two public histopathological datasets for colorectal cancer phenotyping. Our approach achieves consistent performance gains compared to these peer methods.