Scholarlink

<p>The real-world medical datasets are often inherently challenged by imbalanced classes, which impact the performance of deep learning models, leading to overfitting and limited effectiveness. These limitations are particularly pronounced in image segmentation tasks, where accurate delineation of anatomical structures is essential to support clinical decision-making. In order to match the recent advancements and enhance the model’s generalizability and its ability to classify correctly the minor class, specifically the foreground pixels, we applied the generalized dice loss in conjunction with transfer learning, avoiding the redundancy provided by traditional data augmentation techniques and heavy computational data generation strategies. In this paper, we demonstrated that the choice of the loss function plays a pivotal role in optimizing the learning landscape and guiding the model’s training process. The proposed approach generated the highest Dice Coefficient value of 98.44% compared with the existing works and augmentation of 5.24% compared with the network that employed the cross-entropy Loss function. Experimental results indicate that the proposed hybrid approach can accurately identify and segment different shapes of the fetal head, enabling real-time processing and providing a significant potential to assist clinical diagnosis for further circumference measurement.</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

Adaptive transfer learning using SegFormer for imbalanced pixel in medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年5月27日

查看全文
  • Niama Assia El Joudi,
  • Mohamed Lazaar,
  • François Delmotte,
  • Hamid Allaoui,
  • Oussama Mahboub

摘要

The real-world medical datasets are often inherently challenged by imbalanced classes, which impact the performance of deep learning models, leading to overfitting and limited effectiveness. These limitations are particularly pronounced in image segmentation tasks, where accurate delineation of anatomical structures is essential to support clinical decision-making. In order to match the recent advancements and enhance the model’s generalizability and its ability to classify correctly the minor class, specifically the foreground pixels, we applied the generalized dice loss in conjunction with transfer learning, avoiding the redundancy provided by traditional data augmentation techniques and heavy computational data generation strategies. In this paper, we demonstrated that the choice of the loss function plays a pivotal role in optimizing the learning landscape and guiding the model’s training process. The proposed approach generated the highest Dice Coefficient value of 98.44% compared with the existing works and augmentation of 5.24% compared with the network that employed the cross-entropy Loss function. Experimental results indicate that the proposed hybrid approach can accurately identify and segment different shapes of the fetal head, enabling real-time processing and providing a significant potential to assist clinical diagnosis for further circumference measurement.