Scholarlink

<p>Accurate classification of white blood cells plays a crucial role in medical diagnostics, particularly for identifying hematological disorders. However, existing deep learning models often struggle to generalize across complex WBC morphologies due to limitations in local feature extraction and rigid classification heads. This study proposes a hybrid framework that integrates a Vision Transformer (ViT) with Principal Component Analysis (PCA), Neighborhood Components Analysis (NCA), and a fully connected (FC) classifier. The ViT model is fine-tuned using Bayesian Optimization to improve feature representation. Instead of relying solely on the [CLS] token, all class token features are aggregated, decorrelated via PCA, and projected through NCA to enhance class separability. The standard MLP head is replaced by a Bayesian-optimized FC classifier. Experiments on the Raabin-WBC dataset (14,514 images across five classes) show that the proposed method achieves 99.03% accuracy—outperforming the ViT baseline (97.86%) by 1.17%, or 50 additional correctly classified test samples. These results demonstrate that jointly optimizing the ViT architecture and applying supervised dimensionality reduction can significantly boost classification performance in complex biomedical image tasks.</p>

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

PCA-NCA based VIT features for classification of white blood cell images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月27日

查看全文
  • Mohamed Elghandour

摘要

Accurate classification of white blood cells plays a crucial role in medical diagnostics, particularly for identifying hematological disorders. However, existing deep learning models often struggle to generalize across complex WBC morphologies due to limitations in local feature extraction and rigid classification heads. This study proposes a hybrid framework that integrates a Vision Transformer (ViT) with Principal Component Analysis (PCA), Neighborhood Components Analysis (NCA), and a fully connected (FC) classifier. The ViT model is fine-tuned using Bayesian Optimization to improve feature representation. Instead of relying solely on the [CLS] token, all class token features are aggregated, decorrelated via PCA, and projected through NCA to enhance class separability. The standard MLP head is replaced by a Bayesian-optimized FC classifier. Experiments on the Raabin-WBC dataset (14,514 images across five classes) show that the proposed method achieves 99.03% accuracy—outperforming the ViT baseline (97.86%) by 1.17%, or 50 additional correctly classified test samples. These results demonstrate that jointly optimizing the ViT architecture and applying supervised dimensionality reduction can significantly boost classification performance in complex biomedical image tasks.