Scholarlink

<p>The accurate and automated segmentation of brain tumors in MRI images remains one of the most difficult tasks for clinical diagnosis and therapy planning. Segmentation has been significantly improved by recent developments in convolutional neural networks, especially U-Net with skip connections. However, further improvements in accuracy and generalization are still needed. In this study, we propose an integrated attention and residual block U-Net model (ARU-Net) to address the challenges. By employing attention mechanisms, ARU-Net can highlight important areas of processed MRI images while using residual blocks to ensure that the output from deeper networks does not affect the model’s overall performance. Publicly available datasets (TCIA, SciDB and Kaggle) validated the model, measuring the accuracy, loss, and Dice score. The ARU-Net model has performed well, obtaining a maximum training accuracy of 99.41% and validation accuracy of 99.29%. The lowest recorded values of model training and validation losses were 0.48% and 1.04%, respectively. In the last evaluation stage, the model achieved a test accuracy of 99.15%, while the test loss was recorded as 1.91% and achieved a test Dice score of 87.40%. ARU-Net remains a simple and effective way to change U-Net’s residual attention that makes in-domain segmentation better without adding too many parameters. However, the biggest problem is still generalizing across datasets. To further enhance the model’s real-world impact, future research will seek to broaden its application to various medical imaging tasks in clinical contexts. The Codes and models of ARU-Net are available at <a href="https://github.com/Surajit14302/attention_residual_unet_Brain-Tunor_sugmentation">github repository</a>.</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

ARU-Net: a U-Net variant with attention and residual blocks for automated MRI brain tumor segmentation

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Surajit Das,
  • Rajat Subhra Goswami

摘要

The accurate and automated segmentation of brain tumors in MRI images remains one of the most difficult tasks for clinical diagnosis and therapy planning. Segmentation has been significantly improved by recent developments in convolutional neural networks, especially U-Net with skip connections. However, further improvements in accuracy and generalization are still needed. In this study, we propose an integrated attention and residual block U-Net model (ARU-Net) to address the challenges. By employing attention mechanisms, ARU-Net can highlight important areas of processed MRI images while using residual blocks to ensure that the output from deeper networks does not affect the model’s overall performance. Publicly available datasets (TCIA, SciDB and Kaggle) validated the model, measuring the accuracy, loss, and Dice score. The ARU-Net model has performed well, obtaining a maximum training accuracy of 99.41% and validation accuracy of 99.29%. The lowest recorded values of model training and validation losses were 0.48% and 1.04%, respectively. In the last evaluation stage, the model achieved a test accuracy of 99.15%, while the test loss was recorded as 1.91% and achieved a test Dice score of 87.40%. ARU-Net remains a simple and effective way to change U-Net’s residual attention that makes in-domain segmentation better without adding too many parameters. However, the biggest problem is still generalizing across datasets. To further enhance the model’s real-world impact, future research will seek to broaden its application to various medical imaging tasks in clinical contexts. The Codes and models of ARU-Net are available at github repository.