Scholarlink

<p>In medical imaging, deep supervised learning&#xa0;(SL) has seen considerable success but is constrained by its reliance on annotated data. Self-supervised learning (SSL) presents a compelling alternative by utilizing the internal structure of medical images for representation learning without massive amounts of manual labeling. SSL has demonstrated potential to enhance diagnostic accuracy and robustness, addressing critical challenges in traditional supervised methods. In this study, we propose two classification approaches: first, a multiclass classification of a brain MRI dataset with four classes, and second, a multiclass classification through the heterogeneous integration of brain MRI and TB X-ray datasets. Both approaches employ SSL and supervised learning with varying data training percentages (5%, 15%, 25%, 50% and 100%). The Sharpness-Aware Minimization optimizer is integrated into the SSL training routine to improve model generalization by reducing the sharpness of the loss landscape. Our comparative analysis shows that SSL consistently outperforms SL in classification performance across both approaches. These findings underscore SSL's efficacy in improving medical imaging analysis, highlighting its relevance and superiority over traditional supervised techniques.</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

Multiclass classification in medical imaging using self-supervised learning vs supervised learning

  • 期刊论文

  • 发表日期：

    2025年11月29日

查看全文
  • Nitu Kumari,
  • Sonali Agarwal

摘要

In medical imaging, deep supervised learning (SL) has seen considerable success but is constrained by its reliance on annotated data. Self-supervised learning (SSL) presents a compelling alternative by utilizing the internal structure of medical images for representation learning without massive amounts of manual labeling. SSL has demonstrated potential to enhance diagnostic accuracy and robustness, addressing critical challenges in traditional supervised methods. In this study, we propose two classification approaches: first, a multiclass classification of a brain MRI dataset with four classes, and second, a multiclass classification through the heterogeneous integration of brain MRI and TB X-ray datasets. Both approaches employ SSL and supervised learning with varying data training percentages (5%, 15%, 25%, 50% and 100%). The Sharpness-Aware Minimization optimizer is integrated into the SSL training routine to improve model generalization by reducing the sharpness of the loss landscape. Our comparative analysis shows that SSL consistently outperforms SL in classification performance across both approaches. These findings underscore SSL's efficacy in improving medical imaging analysis, highlighting its relevance and superiority over traditional supervised techniques.