ConSlide: Asynchronous Hierarchical Interaction Transformer with Breakup-Reorganize Rehearsal for Continual Whole Slide Image Analysis

• DOI: 10.48550/arxiv.2308.13324
• Publication Date: 2023-01-01
• Authors: Yanyan Huang; Weiqin Zhao; Shujun Wang; Yu Fu; Yuming Jiang; Lequan Yu
• Keywords: Asynchronous learning

Whole slide image (WSI) analysis has become increasingly important in the medical imaging community, enabling automated and objective diagnosis, prognosis, therapeutic-response prediction. However, clinical practice, ever-evolving environment hamper utility of WSI models. In this paper, we propose FIRST continual learning framework for analysis, named ConSlide, to tackle challenges enormous size, utilization hierarchical structure, catastrophic forgetting by progressive model updating on multiple sequential datasets. Our contains three key components. The Hierarchical Interaction Transformer (HIT) is proposed utilize structural knowledge WSI. Breakup-Reorganize (BuRo) rehearsal method developed data replay with efficient region storing buffer reorganizing operation. asynchronous mechanism devised encourage network learn generic specific respectively during stage, based a nested cross-scale similarity (CSSL) module. We evaluated ConSlide four public datasets from TCGA projects. It performs best over other state-of-the-art methods fair WSI-based setting achieves better trade-off overall performance previous task