Scholarlink

<p>Optimization plays an important role in the convergence and generalization of deep learning-based medical image segmentation. Here, we present NoisyAdamW, a novel variant of AdamW that injects adaptive Gaussian noise into gradient updates. The noise amplitude is scaled by the gradient magnitude to preserve stability during late training while encouraging broader loss-surface exploration early on. NoisyAdamW is evaluated on the EffB0-UNet architecture, which uses EfficientNetB0 as the encoder and UNet as the decoder, a strong and efficient baseline for LGG segmentation. Experiments are conducted on the TCGA-LGG dataset and include comparisons with the Adam and AdamW optimizers, as well as a noise hyperparameter analysis. The optimal configuration uses a noise factor of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1 \times 10^{-5}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1</mn> <mo>×</mo> <msup> <mn>10</mn> <mrow> <mo>-</mo> <mn>5</mn> </mrow> </msup> </mrow> </math></EquationSource> </InlineEquation> and achieves a Dice similarity coefficient (DSC) of 0.9388 and an intersection-over-union (IoU) of 0.9117. The results suggest that controlled noise injection can support improved loss-surface exploration and may contribute to better generalization performance. In summary, adaptive noise-based optimization methods, such as NoisyAdamW, show potential for improving CNN-based medical image segmentation and deserve further development and validation for clinical applications.</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

NoisyAdamW: A novel optimizer for EfficientNetB0-UNet (EffB0-UNet) in low-grade glioma segmentation

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Julfa Muhammad Amda,
  • Arya Adhyaksa Waskita,
  • Ade Saputra,
  • Niken Siwi Pamungkas,
  • Zico Pratama Putra,
  • Dede Sutarya,
  • Sita Gandes Pinasti,
  • M. Arif Efendi,
  • Fitrotun Aliyah,
  • Muhammad Yusuf,
  • Dwi Seno Kuncoro Sihono

摘要

Optimization plays an important role in the convergence and generalization of deep learning-based medical image segmentation. Here, we present NoisyAdamW, a novel variant of AdamW that injects adaptive Gaussian noise into gradient updates. The noise amplitude is scaled by the gradient magnitude to preserve stability during late training while encouraging broader loss-surface exploration early on. NoisyAdamW is evaluated on the EffB0-UNet architecture, which uses EfficientNetB0 as the encoder and UNet as the decoder, a strong and efficient baseline for LGG segmentation. Experiments are conducted on the TCGA-LGG dataset and include comparisons with the Adam and AdamW optimizers, as well as a noise hyperparameter analysis. The optimal configuration uses a noise factor of \(1 \times 10^{-5}\) 1 × 10 - 5 and achieves a Dice similarity coefficient (DSC) of 0.9388 and an intersection-over-union (IoU) of 0.9117. The results suggest that controlled noise injection can support improved loss-surface exploration and may contribute to better generalization performance. In summary, adaptive noise-based optimization methods, such as NoisyAdamW, show potential for improving CNN-based medical image segmentation and deserve further development and validation for clinical applications.