Scholarlink

<p>The persistent rise in skin cancer cases has resulted in a high mortality rate of the affected patients due to late detection. In this paper, we proposed a sophisticated deep-learning model MSGNet for the automatic multiclass categorization of skin lesions to address the issues in the multiclass dataset. This work is subdivided into three components. First, in the preprocessing layer, a modified version of SMOTE-ENN is proposed, which uses radius neighbor-based grouping of class elements, an improved data augmentation technique for effectively detecting minority classes in a multiclass environment. Second, in the feature extraction layer, we created a modified VGG19 model which contains parallel Gabor-based convolutional layers in the initial layers. The augmented data is then fed to the parallel convolution layers which capture various extract discriminative features from the same input image. In the third step, in the classification layer, the last layer of modified VGG19 is replaced with two more fully connected layers and one softmax layer to classify skin lesion images. We considered the HAM 10000 and ISIC 2019 datasets in our experiment. MSGNet significantly improved the detection in almost all the classes, and the overall accuracy was found to be 89.33% and 90.62% for the HAM 10000 and ISIC 2019 datasets respectively. Also, we observed significant drift in the F1 score of individual classes which demonstrates that our modified approach efficiently distinguishes each class. It has been experimented that MSGNet performed better than some existing works, and offers a creative and workable way to automate and speed up the task of classifying skin lesions while saving effort, time, and human lives.</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

Multi-class Classification of Class Imbalanced Skin Lesion Dataset Using a Modified SMOTE-ENN Gabor-Enhanced VGG-19 Architecture

  • 期刊论文

  • 发表日期：

    2025年3月8日

查看全文
  • Madhusmita Priyadarshini Sahoo,
  • Rajeswari Sridhar

摘要

The persistent rise in skin cancer cases has resulted in a high mortality rate of the affected patients due to late detection. In this paper, we proposed a sophisticated deep-learning model MSGNet for the automatic multiclass categorization of skin lesions to address the issues in the multiclass dataset. This work is subdivided into three components. First, in the preprocessing layer, a modified version of SMOTE-ENN is proposed, which uses radius neighbor-based grouping of class elements, an improved data augmentation technique for effectively detecting minority classes in a multiclass environment. Second, in the feature extraction layer, we created a modified VGG19 model which contains parallel Gabor-based convolutional layers in the initial layers. The augmented data is then fed to the parallel convolution layers which capture various extract discriminative features from the same input image. In the third step, in the classification layer, the last layer of modified VGG19 is replaced with two more fully connected layers and one softmax layer to classify skin lesion images. We considered the HAM 10000 and ISIC 2019 datasets in our experiment. MSGNet significantly improved the detection in almost all the classes, and the overall accuracy was found to be 89.33% and 90.62% for the HAM 10000 and ISIC 2019 datasets respectively. Also, we observed significant drift in the F1 score of individual classes which demonstrates that our modified approach efficiently distinguishes each class. It has been experimented that MSGNet performed better than some existing works, and offers a creative and workable way to automate and speed up the task of classifying skin lesions while saving effort, time, and human lives.