FMCerNet: frequency-aware multi-label learning for fine-grained cervical cytology screening
TL;DR
The Frequency-Aware Multi-Label Learning Network (FMCerNet), a tile-level feature extractor equipped with a Large-Field Frequency-Enhanced ViT backbone and a Coarse-to-Fine multi-label classification head that improves the accuracy of positive screening and fine-grained subtype classification while maintaining high inference speed is proposed.
Abstract
Abstract Motivation Automated cytology screening is essential for early cervical cancer prevention, where gigapixel whole-slide images (WSIs) are typically partitioned into tiles for tile-level feature extraction and subsequent WSI-level diagnosis aggregation. However, existing tile-level methods remain constrained by limited input sizes, which frequently obscures sparse and morphologically subtle abnormal cells. Furthermore, these methods commonly formulate tile-level representation learning as a coarse single-label classification task, thereby failing to capture the complex co-occurrence patterns and morphological heterogeneity of cervical cytological abnormalities. Results We propose the Frequency-Aware Multi-Label Learning Network (FMCerNet), a tile-level feature extractor equipped with a Large-Field Frequency-Enhanced ViT backbone and a Coarse-to-Fine multi-label classification head. The former incorporates a Dual-Tree Complex Wavelet Transform module to efficiently extract frequency-domain features from large-scale tiles, whereas the latter enhances fine-grained tile representation decoding through hierarchical supervision. Experiments on both public and private datasets demonstrate that FMCerNet improves the accuracy of positive screening and fine-grained subtype classification while maintaining high inference speed. Availability and implementation Code at https://github.com/imAzhou/FMCerNet and archived on Zenodo (doi: 10.5281/zenodo.22886105).