Scholarlink

<p>The diagnosis of skin cancer remains a challenging task due to the high visual similarity among various types of skin lesions. Early detection and accurate diagnosis can significantly improve treatment outcomes. In this study, we propose a lightweight deep learning model based on the U-shape architecture, which leverages lesion segmentation information to enhance the performance of skin lesion diagnosis. To ensure efficient information retention and selection during inference, two novel attention modules have been proposed in the model without significantly increasing the number of trainable parameters. Moreover, we introduce four training strategies designed to exploit the relationship between lesion localization and lesion type classification. To assess the performance of our proposed method, we performed a series of thorough experiments using four benchmark datasets: ISIC2016, ISIC2017, PH2, and HAM10000. The results indicate that our method matches or exceeds the performance of several state-of-the-art approaches, despite relying on significantly fewer parameters and not requiring any supplementary data. These findings suggest that segmentation knowledge can be effectively utilized to improve skin lesion diagnosis in a lightweight and efficient manner, highlighting its potential for integration into real-time and resource-constrained medical diagnostic systems.</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

An efficient dermatological lesion diagnosis method based on U-shape architecture

  • 期刊论文

  • 发表日期：

    2025年11月22日

查看全文
  • Lam Thanh Hien,
  • Pham Trung Hieu,
  • Do Nang Toan

摘要

The diagnosis of skin cancer remains a challenging task due to the high visual similarity among various types of skin lesions. Early detection and accurate diagnosis can significantly improve treatment outcomes. In this study, we propose a lightweight deep learning model based on the U-shape architecture, which leverages lesion segmentation information to enhance the performance of skin lesion diagnosis. To ensure efficient information retention and selection during inference, two novel attention modules have been proposed in the model without significantly increasing the number of trainable parameters. Moreover, we introduce four training strategies designed to exploit the relationship between lesion localization and lesion type classification. To assess the performance of our proposed method, we performed a series of thorough experiments using four benchmark datasets: ISIC2016, ISIC2017, PH2, and HAM10000. The results indicate that our method matches or exceeds the performance of several state-of-the-art approaches, despite relying on significantly fewer parameters and not requiring any supplementary data. These findings suggest that segmentation knowledge can be effectively utilized to improve skin lesion diagnosis in a lightweight and efficient manner, highlighting its potential for integration into real-time and resource-constrained medical diagnostic systems.