Scholarlink

<p>Skin cancer is recognized as one of the most hazardous types of cancer. The physical analysis for the identification of skin cancer is an exhaustive approach that cannot be easily performed by any individual with lesser skill. Therefore, an efficient classifier is needed that can help diagnose skin cancer at an early stage. The presence of noise in dermoscopic images, such as hair, air bubbles, etc., hinders the classification of skin cancer. To eliminate hair and other artifacts from dermoscopic images, this study uses several image preprocessing techniques. In this research, we proposed a novel SkinDWNet (Skin Depthwise Dilated Convolutions Network) based on deep learning for the multiclassification of skin cancer. The SkinDWNet is constructed by using depth-wise dilated convolutions (DDCs) and feature reuse residual blocks (FRBs). The DDCs and FRBs are used to extract relevant features from the dermoscopy images. Furthermore, the Gradient Boosting (GB) technique is employed to identify the prompt feature maps provided by SkinDWNet. SkinDWNet and GB together help to achieve satisfactory improvements in performance. Moreover, the SMOTE Tomek approach is employed to handle the problem of unequal class distribution in the ISIC 2019 data. To assess the performance of the SkinDWNet + GB model, the model is compared with six baseline deep learning models and state-of-the-art (SOTA) models. The SkinDWNet + GB model surpassed the baseline model&#xa0;and achieved the highest&#xa0;accuracy rate of 97.04% on the ISIC 2019 dataset for multiclass classification of skin cancer. Furthermore, ANOVA and McNemar's statistical test demonstrate that the SkinDWNet + GB model outperforms the other models. Thus, the study concludes that the proposed SkinDWNet + GB model produced significant outcomes as compared to baseline models and SOTA models.</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

SkinDWNet: a novel deep learning model for multiclass classification of skin cancers using dermoscopic images

  • 期刊论文

  • 发表日期：

    2025年8月13日

查看全文
  • Ahmad Naeem,
  • Hassaan Malik,
  • Mui-zzud Din,
  • Abolghasem Sadeghi-Niaraki,
  • Daesik Jeong,
  • Rizwan Ali Naqvi

摘要

Skin cancer is recognized as one of the most hazardous types of cancer. The physical analysis for the identification of skin cancer is an exhaustive approach that cannot be easily performed by any individual with lesser skill. Therefore, an efficient classifier is needed that can help diagnose skin cancer at an early stage. The presence of noise in dermoscopic images, such as hair, air bubbles, etc., hinders the classification of skin cancer. To eliminate hair and other artifacts from dermoscopic images, this study uses several image preprocessing techniques. In this research, we proposed a novel SkinDWNet (Skin Depthwise Dilated Convolutions Network) based on deep learning for the multiclassification of skin cancer. The SkinDWNet is constructed by using depth-wise dilated convolutions (DDCs) and feature reuse residual blocks (FRBs). The DDCs and FRBs are used to extract relevant features from the dermoscopy images. Furthermore, the Gradient Boosting (GB) technique is employed to identify the prompt feature maps provided by SkinDWNet. SkinDWNet and GB together help to achieve satisfactory improvements in performance. Moreover, the SMOTE Tomek approach is employed to handle the problem of unequal class distribution in the ISIC 2019 data. To assess the performance of the SkinDWNet + GB model, the model is compared with six baseline deep learning models and state-of-the-art (SOTA) models. The SkinDWNet + GB model surpassed the baseline model and achieved the highest accuracy rate of 97.04% on the ISIC 2019 dataset for multiclass classification of skin cancer. Furthermore, ANOVA and McNemar's statistical test demonstrate that the SkinDWNet + GB model outperforms the other models. Thus, the study concludes that the proposed SkinDWNet + GB model produced significant outcomes as compared to baseline models and SOTA models.