PAGET: Hierarchical Multi-Teacher Knowledge Distillation for Comprehensive Tumor Microenvironment Segmentation

Daisuke Komura, Maki Takao, Mieko Ochi, Takumi Onoyama, Hiroto Katoh, Hiroyuki Abe, Hiroyuki Sano, Teppei Konishi, Toshio Kumasaka, Tomoyuki Yokose, Yohei Miyagi, Tetsuo Ushiku, Shumpei Ishikawa

Keywords: Knowledge distillation, multi-teacher learning, tumor microenvironment, histopathology, semantic segmentation, panoptic segmentation

Abstract: Comprehensive characterization of the tumor microenvironment (TME) from H\&E-stained histopathology images remains challenging due to the diversity of cellular components and limitations of current segmentation methods. We present PAGET (Pathological image segmentation via AGgrEgated Teachers), a multi-teacher knowledge distillation framework that enables simultaneous segmentation of 13 TME components from a single efficient model. Our key insight is that teacher predictions should be aggregated following the biological taxonomy of cell types—from tissue-level context through major cell categories to specific subtypes—rather than simple voting. By training specialized teachers on immunohistochemical restaining data and distilling their aggregated knowledge, the resulting student model not only matches but consistently outperforms the teacher ensemble on external datasets. We provide two complementary variants: PAGET-S for rapid semantic segmentation and PAGET-H for detailed panoptic segmentation. Extensive evaluation across three external datasets demonstrates robust generalization. Our implementation is available at https://github.com/dakomura/PAGET.

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Histopathology

Registration Requirement: Yes

Reproducibility: https://github.com/dakomura/PAGET

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 24