Scholarlink

Data augmentation is essential to improving the efficacy of deep learning models, especially in medical imaging. The wide range of augmentation strategies complicates the identification of the most effective methods for activities. This problem hinders endeavours to enhance model precision and generalisation in skin lesion categorisation. To address this issue, we investigated and contrasted four augmentation techniques: colour space transformations (saturation, hue, brightness, and contrast), noise injection, random erasure, and geometric modifications (rotation, flipping, scaling, cropping, and translation) utilising the HAM10000 dataset. These strategies were assessed across multiple deep learning architectures, with outcomes measured using measures like accuracy, precision, recall, and F1-score. Our findings highlight the important role of specific to the task augmentation. Saturation and colour modifications yielded optimal outcomes, with GoogLeNet attaining the greatest test accuracy of 82.73% and an F1-score of 0.8205. In contrast, techniques such as brightness-contrast adjustments and geometric alterations shown inconsistent efficacy based on the employed architecture. This study shows how carefully selected augmentation procedures can substantially enhance model performance in medical imaging, offering practical guidance for researchers engaged in this intricate yet essential process.

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

Evaluating Augmentation Strategies for CNN-Based Skin Lesion Classification on HAM10000

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rayan Bribesh,
  • Wadhah Zeyad Tareq

摘要

Data augmentation is essential to improving the efficacy of deep learning models, especially in medical imaging. The wide range of augmentation strategies complicates the identification of the most effective methods for activities. This problem hinders endeavours to enhance model precision and generalisation in skin lesion categorisation. To address this issue, we investigated and contrasted four augmentation techniques: colour space transformations (saturation, hue, brightness, and contrast), noise injection, random erasure, and geometric modifications (rotation, flipping, scaling, cropping, and translation) utilising the HAM10000 dataset. These strategies were assessed across multiple deep learning architectures, with outcomes measured using measures like accuracy, precision, recall, and F1-score. Our findings highlight the important role of specific to the task augmentation. Saturation and colour modifications yielded optimal outcomes, with GoogLeNet attaining the greatest test accuracy of 82.73% and an F1-score of 0.8205. In contrast, techniques such as brightness-contrast adjustments and geometric alterations shown inconsistent efficacy based on the employed architecture. This study shows how carefully selected augmentation procedures can substantially enhance model performance in medical imaging, offering practical guidance for researchers engaged in this intricate yet essential process.