Scholarlink

<p>This study proposes a hybrid framework named BlockNet for robust skin cancer classification. Deep features are extracted from intermediate layers of a pre-trained VGG16 network and filtered using an entropy threshold to retain informative representations, which are then fused with handcrafted descriptors including HSV histograms, GLCM texture, and fractal features. A lightweight multi-layer perceptron (MLP) classifier combined with a feature-space balancing strategy is used to address class imbalance and enable efficient classification. Experiments conducted on the VNCancer and HAM10000 datasets demonstrate that BlockNet consistently outperforms state-of-the-art baselines, such as DenseNet121, Xception, MobileNetV2, and InceptionV3. On VNCancer, BlockNet achieves 95.89% accuracy and 95.91% F1-score, significantly higher than the best baseline Xception with 88.58% accuracy and 88.81% F1-score (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p &lt; 0.05\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo>&lt;</mo> <mn>0.05</mn> </mrow> </math></EquationSource> </InlineEquation>). On HAM10000, BlockNet reaches 95.94% accuracy and 95.98% F1-score, substantially outperforming DenseNet121 with 67.71% accuracy and 69.18% F1-score. Statistical analysis using paired t tests confirms the significance of improvements across both datasets (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(p &lt; 0.05\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo>&lt;</mo> <mn>0.05</mn> </mrow> </math></EquationSource> </InlineEquation>). Ablation experiments further show that removing entropy filtering or handcrafted features reduces accuracy by up to 4.3%, emphasizing their complementary contributions. These findings demonstrate that entropy-guided fusion of deep and handcrafted features enhances diagnostic accuracy, improves generalization, and provides a lightweight and practical solution for computer-aided skin cancer diagnosis.</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

Entropy-based feature fusion via BlockNet for robust skin cancer classification

  • 期刊论文

  • 发表日期：

    2025年10月7日

查看全文
  • Thi Trang Nguyen,
  • Van Hieu Vu

摘要

This study proposes a hybrid framework named BlockNet for robust skin cancer classification. Deep features are extracted from intermediate layers of a pre-trained VGG16 network and filtered using an entropy threshold to retain informative representations, which are then fused with handcrafted descriptors including HSV histograms, GLCM texture, and fractal features. A lightweight multi-layer perceptron (MLP) classifier combined with a feature-space balancing strategy is used to address class imbalance and enable efficient classification. Experiments conducted on the VNCancer and HAM10000 datasets demonstrate that BlockNet consistently outperforms state-of-the-art baselines, such as DenseNet121, Xception, MobileNetV2, and InceptionV3. On VNCancer, BlockNet achieves 95.89% accuracy and 95.91% F1-score, significantly higher than the best baseline Xception with 88.58% accuracy and 88.81% F1-score ( \(p < 0.05\) p < 0.05 ). On HAM10000, BlockNet reaches 95.94% accuracy and 95.98% F1-score, substantially outperforming DenseNet121 with 67.71% accuracy and 69.18% F1-score. Statistical analysis using paired t tests confirms the significance of improvements across both datasets ( \(p < 0.05\) p < 0.05 ). Ablation experiments further show that removing entropy filtering or handcrafted features reduces accuracy by up to 4.3%, emphasizing their complementary contributions. These findings demonstrate that entropy-guided fusion of deep and handcrafted features enhances diagnostic accuracy, improves generalization, and provides a lightweight and practical solution for computer-aided skin cancer diagnosis.