DACAS: Distortion-Aware Capacity Allocation for Style-augmented domain generalization in computational pathology

Sun, Shichang, Yang, Haolin, Wang, Haibo, Du, Shun, Liu, Shuang and Lai, Yukun ORCID: https://orcid.org/0000-0002-2094-5680 2026. DACAS: Distortion-Aware Capacity Allocation for Style-augmented domain generalization in computational pathology. Presented at: ICANN 2026, Padua, Italy, 14-17 September 2026. Published in: Pasa, L., Lintas, A., Tetko, I.V., Micheli, A., Navarin, N., Villa, A.E.P., Tortorella, D. and Polato, M. eds. Artificial Neural Networks and Machine Learning – ICANN 2026. Lecture Notes in Computer Science (17092) Springer Nature, pp. 583-594. 10.1007/978-3-032-38407-2_46

Full text not available from this repository.

Abstract

Data augmentation improves domain generalization by increasing distributional coverage, but may also introduce structured distortion that degrades semantic fidelity, especially in structure-sensitive tasks such as computational pathology. Existing approaches primarily address this trade-off by optimizing augmentation policies or softening supervision, implicitly treating all augmentation-induced variations as equally important in the learned representation. In this work, we revisit this problem from a representation perspective. We argue that augmentation reshapes the composition of information, requiring the representation to selectively encode these variations. As a result, the key challenge is not only which variations to generate, but which variations should be preferentially encoded and how representational resources should be allocated accordingly. To this end, we propose a distortion-aware capacity allocation framework that operates after augmentation. We introduce an adaptive discrete bottleneck to adapt representation granularity at the sample level, together with a region-aware distortion variable that modulates spatial encoding according to structural sensitivity. In this way, the model redistributes limited representational resources to preserve structure-critical information while absorbing less informative variations. Experiments on colorectal cancer histopathology show that the proposed approach consistently improves generalization under identical augmentation settings. These results highlight post-augmentation representation allocation as a key mechanism for handling heterogeneous distortion in domain generalization.

Item Type:	Conference or Workshop Item - published (Paper)
Date Type:	Publication
Status:	Published
Schools:	Schools > Computational & Mathematical Sciences Schools > Computer Science & Informatics
Publisher:	Springer Nature
ISBN:	9783032384065
Last Modified:	21 Sep 2026 14:09
URI:	https://orca.cardiff.ac.uk/id/eprint/189741

Actions (repository staff only)

Edit Item