FMCerNet: Frequency-Aware Multi-Label Learning for Fine-Grained Cervical Cytology Screening.

FMCerNet improves cervical cytology screening accuracy by 5.7% in positive detection and enhances subtype classification in gigapixel WSIs using frequency-aware multi-label learning.

• Why it matters: Accurate detection of subtle and sparse abnormal cells in large-scale images is crucial for early cervical cancer diagnosis, but current methods struggle with limited input sizes and complex abnormal patterns.
• What they did: FMCerNet employs a Large-Field Frequency-Enhanced ViT backbone with a Dual-Tree Complex Wavelet Transform to extract frequency features from large tiles and a hierarchical multi-label classification head, tested on public and private datasets.
• The result: The approach boosts positive screening accuracy and subtype classification while maintaining high inference speed, supporting more precise and efficient cervical cytology screening.