Scholarlink

Accurate and efficient Pap-smear classification is critical for early cervical cancer detection. We present an end-to-end pipeline integrating: (1) a GA-tuned DnCNN for noise suppression, (2) Macenko color normalization to harmonize stain variability, (3) data augmentation with SMOTE balancing to address class imbalance, (4) VGG16-BN feature extraction augmented by a Squeeze–Excitation (SE) block for channel-wise attention, and (5) a compact Vision Transformer (ViT) head for final classification. On the seven-class Herlev benchmark, our GA-DnCNN achieves PSNR 43.29 dB and SSIM 0.962; the full VGG16–SE–ViT cascade converges in ≈12 epochs to 97.1 \% overall accuracy (± 0.4 \%) and 96.5 \% macro-F\(_1\) (±0.5 \%). With only ≈34 M parameters and under two hours of training on a single NVIDIA P100 GPU, our method matches or exceeds the performance of much larger models.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient Pap-Smear Classification via GA-DnCNN Denoising and SE-Gate-Guided VGG16-ViT

  • 章节

  • 发表日期：

    2026年

查看全文
  • Salma Oussahi,
  • Aziz Darouichi,
  • El Mahdi El Guarmah

摘要

Accurate and efficient Pap-smear classification is critical for early cervical cancer detection. We present an end-to-end pipeline integrating: (1) a GA-tuned DnCNN for noise suppression, (2) Macenko color normalization to harmonize stain variability, (3) data augmentation with SMOTE balancing to address class imbalance, (4) VGG16-BN feature extraction augmented by a Squeeze–Excitation (SE) block for channel-wise attention, and (5) a compact Vision Transformer (ViT) head for final classification. On the seven-class Herlev benchmark, our GA-DnCNN achieves PSNR 43.29 dB and SSIM 0.962; the full VGG16–SE–ViT cascade converges in ≈12 epochs to 97.1 \% overall accuracy (± 0.4 \%) and 96.5 \% macro-F\(_1\) (±0.5 \%). With only ≈34 M parameters and under two hours of training on a single NVIDIA P100 GPU, our method matches or exceeds the performance of much larger models.