Scholarlink

<p>Brain tumor segmentation is necessary for both early tumor diagnosis and radiotherapy planning. The intricate aspects of brain tumor MRI images, such as significant tumor appearance variability and ambiguous tumor boundaries, make it challenging to improve tumor segmentation approaches. Numerous mathematical techniques have been developed till date with a focus on reducing the number of parameters in the neural network model. This work uses a novel weighted standout gaussian dropout mechanism incorporated in the bottleneck of the encoder-decoder architecture. The trade-off is maintained using the incorporation of the fuzzy attention mechanism in the dense module of the architecture. The significance of using the standout gaussian dropout over the standard dropout can be stated from the fact that it allows smooth gradient flow, which does not eliminate the complete weights thereby keeping the accuracy up to the mark with the reduction in the number of parameters. The incorporation of fuzzy attention mechanism in the dense feature fusion block (DFFB) and skip connection relocation enhances diagnostic precision. The input feature variability is increased with progressive enhancement module (PEM), which makes the model learn on varying features making it more robust. The number of parameters has been reduced from the baseline U-Net from 25,526,626 to 4,510,882 in the proposed model. Along with that, the parameters of the proposed model have been reduced from 97.38&#xa0;MB in baseline U-Net to 17.21&#xa0;MB in the proposed architecture. The average dice similarity of the proposed method for segmenting complete tumor is 0.8582. The loss is reduced to 0.02207. These results are comparable to other state-of-the-art methods.</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

Feature diversity and stochastic regularization based generalized brain tumor segmentation

  • 期刊论文

  • 发表日期：

    2025年9月19日

查看全文
  • Poonam Rani Verma,
  • Ashish Kumar Bhandari

摘要

Brain tumor segmentation is necessary for both early tumor diagnosis and radiotherapy planning. The intricate aspects of brain tumor MRI images, such as significant tumor appearance variability and ambiguous tumor boundaries, make it challenging to improve tumor segmentation approaches. Numerous mathematical techniques have been developed till date with a focus on reducing the number of parameters in the neural network model. This work uses a novel weighted standout gaussian dropout mechanism incorporated in the bottleneck of the encoder-decoder architecture. The trade-off is maintained using the incorporation of the fuzzy attention mechanism in the dense module of the architecture. The significance of using the standout gaussian dropout over the standard dropout can be stated from the fact that it allows smooth gradient flow, which does not eliminate the complete weights thereby keeping the accuracy up to the mark with the reduction in the number of parameters. The incorporation of fuzzy attention mechanism in the dense feature fusion block (DFFB) and skip connection relocation enhances diagnostic precision. The input feature variability is increased with progressive enhancement module (PEM), which makes the model learn on varying features making it more robust. The number of parameters has been reduced from the baseline U-Net from 25,526,626 to 4,510,882 in the proposed model. Along with that, the parameters of the proposed model have been reduced from 97.38 MB in baseline U-Net to 17.21 MB in the proposed architecture. The average dice similarity of the proposed method for segmenting complete tumor is 0.8582. The loss is reduced to 0.02207. These results are comparable to other state-of-the-art methods.