Scholarlink

<p>Automated classification of skin diseases using deep learning and machine learning is becoming increasingly prominent in the field of medical imaging. This study explores a hybrid framework that combines convolutional neural networks (CNNs), vision transformers (ViTs), and traditional machine learning classifiers to distinguish between three common skin conditions: Psoriasis, Eczema, and Tinea Ringworm. A curated dataset of 2,750 clinical images sourced from DermNet and the Atlas of Dermatology was used, with two dataset splits: 70:15:15 and 80:10:10 for training, validation, and testing, respectively. Feature extraction was performed using five deep learning backbones—EfficientNet-B0, ResNet50, ConvNeXt-Tiny, Swin-Tiny, and DeiT-Small. To reduce dimensionality and enhance relevance, five feature selection techniques were applied: Boruta, Chi-Square, ANOVA, recursive feature elimination (RFE), and Kruskal–Wallis (KW). The selected features were then used to train five classical machine learning classifiers: support vector machine (SVM), logistic regression (LR), k-nearest neighbors (KNN), random forest (RF), and multilayer perceptron (MLP). All experiments were repeated with three random seeds (42, 99, and 123) to ensure robustness. Among the combinations, ConvNeXt-Tiny features selected via Boruta and classified using SVM delivered the best hybrid result, achieving an accuracy of 79.78%, a notable improvement over the baseline of 74.7% ± 0.0% on the 80:10:10 split. Additionally, an end-to-end classification approach was evaluated using four backbone models—ResNet50, EfficientNet-B0, ConvNeXt-Tiny, and Swin-Tiny—also across three runs with seeds 42, 99, and 123. Among these, the Swin-Tiny model achieved the highest performance, with an accuracy of 82.1% ± 1.3%. These findings demonstrate the strength of hybrid approaches that leverage both deep learning for feature extraction and classical machine learning for classification, alongside the competitive edge of end-to-end deep learning, particularly in dermatological image analysis.</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

A hybrid CNN–ViT framework for skin disease classification via feature extraction and selection

  • 期刊论文

  • 发表日期：

    2025年9月18日

查看全文
  • Pragya Gupta,
  • Nimish Vadgaonkar,
  • Jagannath Nirmal,
  • Ninad Mehendale

摘要

Automated classification of skin diseases using deep learning and machine learning is becoming increasingly prominent in the field of medical imaging. This study explores a hybrid framework that combines convolutional neural networks (CNNs), vision transformers (ViTs), and traditional machine learning classifiers to distinguish between three common skin conditions: Psoriasis, Eczema, and Tinea Ringworm. A curated dataset of 2,750 clinical images sourced from DermNet and the Atlas of Dermatology was used, with two dataset splits: 70:15:15 and 80:10:10 for training, validation, and testing, respectively. Feature extraction was performed using five deep learning backbones—EfficientNet-B0, ResNet50, ConvNeXt-Tiny, Swin-Tiny, and DeiT-Small. To reduce dimensionality and enhance relevance, five feature selection techniques were applied: Boruta, Chi-Square, ANOVA, recursive feature elimination (RFE), and Kruskal–Wallis (KW). The selected features were then used to train five classical machine learning classifiers: support vector machine (SVM), logistic regression (LR), k-nearest neighbors (KNN), random forest (RF), and multilayer perceptron (MLP). All experiments were repeated with three random seeds (42, 99, and 123) to ensure robustness. Among the combinations, ConvNeXt-Tiny features selected via Boruta and classified using SVM delivered the best hybrid result, achieving an accuracy of 79.78%, a notable improvement over the baseline of 74.7% ± 0.0% on the 80:10:10 split. Additionally, an end-to-end classification approach was evaluated using four backbone models—ResNet50, EfficientNet-B0, ConvNeXt-Tiny, and Swin-Tiny—also across three runs with seeds 42, 99, and 123. Among these, the Swin-Tiny model achieved the highest performance, with an accuracy of 82.1% ± 1.3%. These findings demonstrate the strength of hybrid approaches that leverage both deep learning for feature extraction and classical machine learning for classification, alongside the competitive edge of end-to-end deep learning, particularly in dermatological image analysis.