Scholarlink

Skin disorders encompass a broad spectrum of conditions that alter both the appearance and function of the skin, affecting a vast number of individuals worldwide. Accurate classification of these diseases is essential for timely diagnosis and effective treatment. This study utilizes the HAM10000 dataset, sourced from Kaggle, to evaluate the performance of Transfer Learning architectures, namely InceptionV3, ResNet50, and DenseNet121. Due to the imbalances in the dataset, models are trained with and without methods to mitigate class distribution issues. Data augmentation and class weighting strategies are employed to solve the problem. Evaluation metrics include classification accuracy and Area Under the Curve (AUC). Without any balancing methods, DenseNet121 performs best, achieving 0.77 accuracy and an AUC of 0.90. With balancing strategies, InceptionV3 gives the best results with an improved accuracy of 0.77 and an AUC of 0.94. The findings of the research are that handling data imbalance enhances the reliability of automated skin disease diagnosis.

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

Optimizing Skin Disease Classification: A Transfer Learning Framework for Imbalanced Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tamanna,
  • Ritika Kumari

摘要

Skin disorders encompass a broad spectrum of conditions that alter both the appearance and function of the skin, affecting a vast number of individuals worldwide. Accurate classification of these diseases is essential for timely diagnosis and effective treatment. This study utilizes the HAM10000 dataset, sourced from Kaggle, to evaluate the performance of Transfer Learning architectures, namely InceptionV3, ResNet50, and DenseNet121. Due to the imbalances in the dataset, models are trained with and without methods to mitigate class distribution issues. Data augmentation and class weighting strategies are employed to solve the problem. Evaluation metrics include classification accuracy and Area Under the Curve (AUC). Without any balancing methods, DenseNet121 performs best, achieving 0.77 accuracy and an AUC of 0.90. With balancing strategies, InceptionV3 gives the best results with an improved accuracy of 0.77 and an AUC of 0.94. The findings of the research are that handling data imbalance enhances the reliability of automated skin disease diagnosis.