[Paper] TICON: A Slide-Level Tile Contextualizer for Histopathology Representation Learning

Source: arXiv - 2512.21331v1

Overview

The paper introduces TICON, a transformer‑based “tile contextualizer” that enriches the feature vectors of tiny image patches (tiles) extracted from whole‑slide pathology scans. By injecting slide‑level context into any pre‑trained tile encoder, TICON bridges the gap between local (tile‑wise) and global (slide‑wise) analysis, delivering state‑of‑the‑art performance on a suite of computational pathology benchmarks.

Key Contributions

• Universal contextualizer: Works with embeddings from any tile‑level foundation model (e.g., ResNet, ViT, CLIP‑style encoders).
• Masked tile modeling pre‑training: A self‑supervised objective that forces the transformer to predict missing tile embeddings, thereby learning slide‑wide relationships.
• Unified encoder for diverse tasks: A single shared network replaces the need for task‑specific tile encoders, simplifying pipelines.
• Strong empirical gains: Sets new SOTA on tile‑level benchmarks (HEST‑Bench, THUNDER, CATCH) and slide‑level benchmark (Patho‑Bench).
• Efficient slide‑level foundation model: Pre‑trains a slide aggregator on top of TICON using only 11 K WSIs, outperforming models trained on up to 350 K WSIs.

Methodology

1. Tile Embedding Extraction – Existing pathology foundation models generate a raw embedding for each tile (e.g., 256‑dim vector).
2. Contextualizer Architecture – TICON stacks a standard Vision Transformer (ViT) encoder that treats each tile embedding as a token. Positional encodings reflect the tile’s spatial location on the slide.
3. Masked Tile Modeling (MTM) – During pre‑training, a random subset of tile tokens is masked. The model must reconstruct the missing embeddings from the surrounding context, encouraging it to capture slide‑level patterns (tissue architecture, tumor‑stroma interactions, etc.).
4. Fine‑tuning / Aggregation – For downstream tasks, the contextualized tile embeddings are either fed directly to a classifier (tile‑level tasks) or pooled by a lightweight slide‑level aggregator (e.g., a shallow transformer or attention‑based pooling) to produce a slide representation.
5. Plug‑and‑Play Compatibility – Because TICON only consumes embeddings, any new tile encoder can be swapped in without retraining the contextualizer.

Results & Findings

• Data efficiency: The slide‑level aggregator trained on just 11 K WSIs beats competitors that used 30–350 K WSIs.
• Cross‑model robustness: When swapping the underlying tile encoder (ResNet‑50, Swin‑Transformer, CLIP‑Vision), TICON consistently lifts performance, confirming its “any‑encoder” claim.
• Ablation: Removing the MTM objective drops performance by ~3 % on average, highlighting the importance of self‑supervised context learning.

Practical Implications

• Simplified pipelines – Teams can adopt a single TICON service to add context to any tile embeddings, removing the need to maintain multiple task‑specific encoders.
• Faster model iteration – Since only the contextualizer needs fine‑tuning for a new downstream task, developers can experiment with new objectives (e.g., weak supervision, active learning) without re‑training massive tile‑level backbones.
• Reduced data requirements – The slide‑level foundation model achieves SOTA with an order‑of‑magnitude fewer WSIs, lowering storage and annotation costs for hospitals and biotech firms.
• Edge deployment – Tile embeddings can be computed on‑device (e.g., on a GPU‑accelerated scanner), then sent to a lightweight TICON server for contextualization, enabling real‑time assistance in the pathology lab.
• Transferability – Because TICON operates on generic embeddings, it can be repurposed for related domains (e.g., radiology patches, satellite imagery) where local patches need global context.

Limitations & Future Work

• Spatial granularity – TICON treats tiles as a flat token sequence; extremely large slides may still suffer from limited receptive field unless hierarchical tokenization is added.
• Memory footprint – Processing thousands of tiles per slide can be GPU‑intensive; the authors suggest future work on memory‑efficient attention (e.g., Linformer, Performer).
• Domain shift – While robust across tile encoders, the model’s performance on slides from entirely new staining protocols or scanners remains to be evaluated.
• Explainability – The transformer’s attention maps provide some insight, but more interpretable mechanisms (e.g., concept bottlenecks) could help clinicians trust the predictions.

Bottom line: TICON offers a plug‑and‑play, data‑efficient way to inject slide‑level context into any tile representation, delivering measurable gains across a spectrum of pathology tasks. For developers building AI‑assisted pathology tools, it promises a cleaner architecture, lower data barriers, and a path toward more globally aware visual models.

Authors

• Varun Belagali
• Saarthak Kapse
• Pierre Marza
• Srijan Das
• Zilinghan Li
• Sofiène Boutaj
• Pushpak Pati
• Srikar Yellapragada
• Tarak Nath Nandi
• Ravi K Madduri
• Joel Saltz
• Prateek Prasanna
• Stergios Christodoulidis Maria Vakalopoulou
• Dimitris Samaras

Paper Information

• arXiv ID: 2512.21331v1
• Categories: cs.CV
• Published: December 24, 2025
• PDF: Download PDF