# CRC-Former: frequency-domain adaptive swin-transformer for colorectal cancer histopathology classification

## TL;DR

This paper introduces CRC-Former, a new deep learning model that improves colorectal cancer diagnosis by combining frequency analysis with advanced image processing techniques.

## Contribution

The novel CRC-Former architecture integrates frequency-aware learning and cross-scale modeling for improved histopathology classification.

## Key findings

- CRC-Former outperforms existing models on the Chaoyang CRC dataset.
- The model effectively captures both fine-grained textures and global context in histopathology images.

## Abstract

Colorectal cancer (CRC) diagnosis from whole-slide histopathology images remains challenging due to pronounced tissue heterogeneity, multi-scale morphological variations, and the subtle nature of early neoplastic changes. While deep learning models have shown promise, conventional architectures struggle to simultaneously capture fine-grained texture cues and global architectural context, often overlooking diagnostically critical frequency-domain signatures.

To address these limitations, we propose CRC-Former, a novel hybrid architecture that synergistically integrates frequency-aware representation learning with efficient cross-scale sequence modeling. Specifically, CRC-Former introduces two key components: (i) a Frequency-aware Global-Local Transformer Block (FGT), which decomposes features via Haar wavelet transform and applies orientation-specific sliding-window attention in distinct subbands to enhance sensitivity to multi-directional pathological textures; and (ii) a Cross-Scale Mamba Block (CSM), which leverages selective state-space modeling to fuse hierarchical features across resolutions with linear complexity.

Evaluated on the large-scale Chaoyang CRC dataset, CRC-Former achieves state-of-the-art performance, outperforming strong baselines.

Our work demonstrates that explicit integration of signal processing priors with modern sequence modeling offers a powerful paradigm for robust, interpretable, and scalable computational pathology.