Gigapixel pathology reasoning

TC-SSA

Token Compression via Semantic Slot Aggregation compresses variable-length whole-slide image features into a fixed budget of semantic slots for efficient vision-language reasoning.

Zhuo Chen^1,2, Xiaoyu Yang¹, and Lijian Xu^1,*

Paper GitHub

1.7% Visual tokens retained

78.34% SlideBench TCGA overall

77.14% Diagnosis accuracy

98.27% TCGA-NSCLC AUC

79.80% PANDA ISUP accuracy

Authors

Research team

Zhuo Chen^1,2, Xiaoyu Yang¹, and Lijian Xu^1,*

¹ Shenzhen University of Advanced Technology, Shenzhen, Guangdong, China. ² University of Nottingham Ningbo China, FoSE, Ningbo, Zhejiang, China. ^* Corresponding author: xulijian@suat-sz.edu.cn.

Architecture

Semantic aggregation instead of patch pruning

TC-SSA routes WSI patch embeddings through a gated Top-2 assignment module, aggregates them into K semantic slots, and refines each slot independently. The design keeps all slide evidence in the representation while retaining only 1.7% of visual tokens.

TC-SSA architecture and whole-flow diagram — Whole-slide image features are routed into semantic slots with robust semantic slot regularization.

Evaluation

Efficient compression with competitive downstream performance

TC-SSA is evaluated on SlideBench VQA, zero-shot pathology VQA, TCGA MIL classification, and PANDA ISUP grading. It uses 1.72T FLOPs under token compression, compared with 133.3T FLOPs for the uncompressed SlideChat upper bound.

SlideBench and Zero-Shot VQA

Accuracy on TCGA, BCNB, and WSI-VQA* with FLOPs for compressed inference.

Method	FLOPs	Microscopy	Diagnosis	Overall	BCNB	WSI-VQA*
SlideChat	133.3T	87.64	73.27	81.17	54.14	60.18
LLaVA-Med	1.70T	47.34	32.78	42.00	30.10	26.31
Quilt-LLaVA	1.70T	57.76	35.96	48.07	32.19	44.43
MedDr	1.70T	73.30	57.78	67.70	33.67	54.36
TC-SSA	1.72T	81.94	77.14	78.34	55.94	56.62

MIL Classification and PANDA Grading

BRCA and NSCLC use AUC; PANDA reports ISUP grading accuracy.

Encoder	Method	TCGA-BRCA	TCGA-NSCLC	PANDA
GigaPath	RRTMIL	94.82	97.63	72.46
GigaPath	2DMamba	93.84	96.87	75.72
UNI	RRTMIL	94.61	97.88	74.93
UNI	2DMamba	93.08	97.14	76.37
UNI	TC-SSA	95.83	98.27	79.80

Efficiency and effectiveness comparison on SlideBench VQA.

Slot-count ablation chart — Slot-count ablation for TCGA-BRCA and TCGA-NSCLC.

Expert specialization

Slots learn coherent tissue semantics

Multi-dataset t-SNE visualizations show patch embeddings grouped by assigned expert or slot. The resulting clusters indicate that semantic slot aggregation produces consistent tissue-level organization across datasets.

Expert clustering t-SNE visualization — Expert assignment induces coherent, dataset-consistent tissue clusters.

Citation

Paper and code

Read the paper on arXiv or use the GitHub repository for training, evaluation, and VQA experiments.

@article{chen2026tcssa,
  title={TC-SSA: Token Compression via Semantic Slot Aggregation for Gigapixel Pathology Reasoning},
  author={Chen, Zhuo and Yang, Xiaoyu and Xu, Lijian},
  journal={arXiv preprint arXiv:2603.01143},
  year={2026}
}