Scholarlink

<p>Skin cancer is one of the most common and life-threatening diseases in the world, and successful treatments rely on timely and precise detection. Deep learning has achieved remarkable success in medical image analysis, with hybrid models combining multiple techniques showing promising results in enhancing diagnostic accuracy. However, their performance is often compromised due to the challenges inherited in dermatological datasets, such as class imbalance, noisy annotations, and variations in image quality and acquisition conditions. Because of such issues generalization of results becomes limited and it also reduces the reliability of existing models in real-world clinical applications. Therefore, in this research, we propose a robust hybrid framework for skin lesion detection that integrates hybrid deep learning models via transfer learning along with dataset refinement techniques. To improve image quality and highlight lesion features, the model incorporates advanced preprocessing steps, including Z-score standardization, lesion region isolation via segmentation masking, artifact removal, noise reduction, and contrast enhancement using the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. These refinements enhance model learning and improve classification performance. High-level features are extracted using pretrained Convolutional Neural Network (CNN) architectures such as VGG19, MobileNetv2, ResNet50, AlexNet, DenseNet121, and EfficientNetB0, followed by hybrid feature fusion for effective classification. Unlike existing hybrid frameworks that focus primarily on architecture depth, this study introduces a novel integration of quantified dataset refinement (via CLAHE and Dull-Razor) with a dual-branch feature fusion network (DenseNet121 + EfficientNetB0). This approach specifically targets the research gap of improving feature extractability in low contrast dermatoscopic images before classification. Our hybrid framework demonstrated high accuracy (98.89%), precision (0.96%), recall (0.97%), and F1-score (0.96%) across a range of skin lesion types tested on the publicly accessible HAM10000 dataset. These results are comparable to or better than current state-of-the-art methods 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 deep learning based robust framework for enhancing the prediction of skin lesions

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Sheeza Naeem,
  • Farah Haneef,
  • Muhammad Nouman Noor

摘要

Skin cancer is one of the most common and life-threatening diseases in the world, and successful treatments rely on timely and precise detection. Deep learning has achieved remarkable success in medical image analysis, with hybrid models combining multiple techniques showing promising results in enhancing diagnostic accuracy. However, their performance is often compromised due to the challenges inherited in dermatological datasets, such as class imbalance, noisy annotations, and variations in image quality and acquisition conditions. Because of such issues generalization of results becomes limited and it also reduces the reliability of existing models in real-world clinical applications. Therefore, in this research, we propose a robust hybrid framework for skin lesion detection that integrates hybrid deep learning models via transfer learning along with dataset refinement techniques. To improve image quality and highlight lesion features, the model incorporates advanced preprocessing steps, including Z-score standardization, lesion region isolation via segmentation masking, artifact removal, noise reduction, and contrast enhancement using the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. These refinements enhance model learning and improve classification performance. High-level features are extracted using pretrained Convolutional Neural Network (CNN) architectures such as VGG19, MobileNetv2, ResNet50, AlexNet, DenseNet121, and EfficientNetB0, followed by hybrid feature fusion for effective classification. Unlike existing hybrid frameworks that focus primarily on architecture depth, this study introduces a novel integration of quantified dataset refinement (via CLAHE and Dull-Razor) with a dual-branch feature fusion network (DenseNet121 + EfficientNetB0). This approach specifically targets the research gap of improving feature extractability in low contrast dermatoscopic images before classification. Our hybrid framework demonstrated high accuracy (98.89%), precision (0.96%), recall (0.97%), and F1-score (0.96%) across a range of skin lesion types tested on the publicly accessible HAM10000 dataset. These results are comparable to or better than current state-of-the-art methods in dermatological image analysis.