Scholarlink

<p>The segmentation of brain tumors in magnetic resonance images (MRI) poses significant challenges due to inherent uncertainties and complexities, including intra-class variation, inter-class similarity, noise, and artifacts. To address these challenges, this research proposes the Fuzzy Inception UNet model, a hybrid approach combining fuzzy logic with deep learning to surpass existing segmentation methods. Unlike conventional models, the proposed approach incorporates a fuzzy residual convolution block featuring an Adaptive Neuro-Fuzzy Inference System (ANFIS) to more effectively manage uncertainties and noise. Additionally, the inception module with dimensionality reduction (IMDR) captures multi-level contextual features, significantly improving segmentation accuracy. Transfer learning and reinforcement learning techniques are employed to define optimal hyperparameters for the Gaussian membership functions, which play a critical role in the fuzzification layer of ANFIS. Empirical evaluations demonstrate the efficacy of the proposed method, with notable performance metrics. The model achieves a Dice score of 89.66%, specificity of 99.37%, sensitivity of 92.01%, and a Hausdorff distance of 10.88, highlighting its accuracy and robustness in segmenting brain tumors from MRI images. By comparing against existing methods, the proposed approach demonstrates superior performance compared to existing fuzzy logic or deep learning approaches. Comprehensive ablation studies further underscore the importance of the proposed method in enhancing segmentation performance.</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

Fuzzy Inception UNet: Bridging Uncertainties in MRI for Multi-Modal Brain Tumor Segmentation

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Kaliprasad Pani,
  • Indu Chawla

摘要

The segmentation of brain tumors in magnetic resonance images (MRI) poses significant challenges due to inherent uncertainties and complexities, including intra-class variation, inter-class similarity, noise, and artifacts. To address these challenges, this research proposes the Fuzzy Inception UNet model, a hybrid approach combining fuzzy logic with deep learning to surpass existing segmentation methods. Unlike conventional models, the proposed approach incorporates a fuzzy residual convolution block featuring an Adaptive Neuro-Fuzzy Inference System (ANFIS) to more effectively manage uncertainties and noise. Additionally, the inception module with dimensionality reduction (IMDR) captures multi-level contextual features, significantly improving segmentation accuracy. Transfer learning and reinforcement learning techniques are employed to define optimal hyperparameters for the Gaussian membership functions, which play a critical role in the fuzzification layer of ANFIS. Empirical evaluations demonstrate the efficacy of the proposed method, with notable performance metrics. The model achieves a Dice score of 89.66%, specificity of 99.37%, sensitivity of 92.01%, and a Hausdorff distance of 10.88, highlighting its accuracy and robustness in segmenting brain tumors from MRI images. By comparing against existing methods, the proposed approach demonstrates superior performance compared to existing fuzzy logic or deep learning approaches. Comprehensive ablation studies further underscore the importance of the proposed method in enhancing segmentation performance.