Revisiting Adaptive Cellular Recognition Under Domain Shifts: A Contextual Correspondence View

TL;DR

This work tackles cross-domain cellular nuclei recognition under organ and stain shifts by introducing a hierarchical pathology data genesis view and a multifaceted correspondence discovery framework. It leverages surrogate tasks—nuclei masking/restoration for nucleus–tissue and inter-nuclei correspondences—and a self-adaptive dynamic distillation strategy that fuses CNN and ViT representations, guided by instance-wise uncertainty. The approach yields substantial improvements over state-of-the-art UDA methods on cross-organ and cross-stain benchmarks, validating the effectiveness of exploiting high-level biological correspondences rather than relying solely on pixel-level or pseudo-label-driven alignment. Overall, the method provides a robust, domain-invariant mechanism for nuclei recognition with practical implications for generalizable digital pathology workflows.

Abstract

Cellular nuclei recognition serves as a fundamental and essential step in the workflow of digital pathology. However, with disparate source organs and staining procedures among histology image clusters, the scanned tiles inherently conform to a non-uniform data distribution, which induces deteriorated promises for general cross-cohort usages. Despite the latest efforts leveraging domain adaptation to mitigate distributional discrepancy, those methods are subjected to modeling the morphological characteristics of each cell individually, disregarding the hierarchical latent structure and intrinsic contextual correspondences across the tumor micro-environment. In this work, we identify the importance of implicit correspondences across biological contexts for exploiting domain-invariant pathological composition and thereby propose to exploit the dependence over various biological structures for domain adaptive cellular recognition. We discover those high-level correspondences via unsupervised contextual modeling and use them as bridges to facilitate adaptation over diverse organs and stains. In addition, to further exploit the rich spatial contexts embedded amongst nuclear communities, we propose self-adaptive dynamic distillation to secure instance-aware trade-offs across different model constituents. The proposed method is extensively evaluated on a broad spectrum of cross-domain settings under miscellaneous data distribution shifts and outperforms the state-of-the-art methods by a substantial margin. Code is available at https://github.com/camwew/CellularRecognition_DA_CC.

Figures (8)

• Figure 1: (a) Illustrative results for cellular recognition under domain shifts. X$\rightarrow$Y denotes that the model is trained with data from X organ and then evaluated on Y organ. bDice and bPQ measure the accuracy of class-agnostic segmentation. $F^*$ denote the $F$ score for different nuclear types, indicating subtyping accuracy. (b) Schematic diagram of the hierarchical nature of latent variables. The pathological composition principle of tumor micro-environment inherits fundamental invariance regardless of the underlying confounding factors such as sampling organs and staining protocols, holding great promises to formalize domain-agnostic biomarkers.
• Figure 2: Exemplary H&E-stained histology tiles. In each sub-figure, red, yellow, blue, and green rectangles correspond to the nuclei of neoplastic, epithelial, connective, and inflammatory cells, respectively.
• Figure 3: Conceptual illustration of the insight. To exploit the intrinsic pathological composition principle which inherits cross-domain coherence, we propose to devise self-supervised surrogate tasks to discover multifaceted biological correspondences, from which the high-level principle variables can be implicitly learned to endow the model with strengthened generalizability.
• Figure 4: Overview of the proposed approach. We aim to learn the implicit correspondences across various biological structures via self-regulated surrogate tasks. Specifically, we first perform nuclei masking and then learn to restore the obscured contextual details based on the characteristics of tissue and neighbouring nuclei. For correspondence discovery within nuclear cluster, the dotted bounding box and features indicate the location and mask token of the masked nucleus.
• Figure 5: Qualitative comparison of nuclei recognition results. Images in the top two rows are from the testis, whereas images in the third and fourth rows are from the thyroid and bile-duct, respectively. In each sub-figure, red, yellow, blue, and green contours correspond to the nuclei of neoplastic, epithelial, connective, and inflammatory cells, respectively.

Theorems & Definitions (1)

• proposition thmcounterproposition: Hierarchical Formulation of Latent Variables