Scholarlink

<p>Brain tumors are a serious medical condition that requires careful diagnosis and prompt treatment, like gliomas, which are the most common malignant brain tumors of varying degrees that significantly determine the survival rate of patients. The utilization of magnetic resonance imaging (MRI) for the segmentation and categorization of tumors is prevalent and essential for facilitating early diagnosis and treatment planning. To achieve this clinical need, taking into account the limited computational resources in many health centers, we propose in this paper a dual approach of modification to improve the efficiency of model performance and enrich the data both. We adopt transfer learning technology to choose VGG19 as the encoder for the Unet model for its ability to extract features from images because of its advanced structure and its ability to learn complex and hierarchical features. It has been trained on a large dataset such as ImageNet, which makes it able to extract generic and powerful features that can be transferred to other tasks via the learning transfer technology. And the development of the decoding part by replacing regular convolutions with deep ones to reduce the computational cost and maintain the good performance of the model, making the model lighter and faster in training and inference. Detachable deep convolutions work by applying one convolutional filter to each input channel separately and then combining the outputs using 1×1 convolutions. This approach reduces the risk of over-allocation and improves performance in applications running on devices with limited resources. The structure of the model affects not only its performance but also the diversity of the data on which it is trained by applying the Laplace filter to the data and taking advantage of its ability to extract attributes from images such as edges and fine details and integrate them into the original images to preserve the original attributes without losing them. The results confirm the effectiveness of the approach by achieving a better accuracy of 99.49%, a dice similarity coefficient (DSC) of 91.28%, and Intersection over Onion (IoU) of 91.70%.</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

Improved brain tumor segmentation using a modified VGG-Unet with depth-wise convolutions in decoder and Laplacian filter

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Lahmar Hanine,
  • Naimi Hilal

摘要

Brain tumors are a serious medical condition that requires careful diagnosis and prompt treatment, like gliomas, which are the most common malignant brain tumors of varying degrees that significantly determine the survival rate of patients. The utilization of magnetic resonance imaging (MRI) for the segmentation and categorization of tumors is prevalent and essential for facilitating early diagnosis and treatment planning. To achieve this clinical need, taking into account the limited computational resources in many health centers, we propose in this paper a dual approach of modification to improve the efficiency of model performance and enrich the data both. We adopt transfer learning technology to choose VGG19 as the encoder for the Unet model for its ability to extract features from images because of its advanced structure and its ability to learn complex and hierarchical features. It has been trained on a large dataset such as ImageNet, which makes it able to extract generic and powerful features that can be transferred to other tasks via the learning transfer technology. And the development of the decoding part by replacing regular convolutions with deep ones to reduce the computational cost and maintain the good performance of the model, making the model lighter and faster in training and inference. Detachable deep convolutions work by applying one convolutional filter to each input channel separately and then combining the outputs using 1×1 convolutions. This approach reduces the risk of over-allocation and improves performance in applications running on devices with limited resources. The structure of the model affects not only its performance but also the diversity of the data on which it is trained by applying the Laplace filter to the data and taking advantage of its ability to extract attributes from images such as edges and fine details and integrate them into the original images to preserve the original attributes without losing them. The results confirm the effectiveness of the approach by achieving a better accuracy of 99.49%, a dice similarity coefficient (DSC) of 91.28%, and Intersection over Onion (IoU) of 91.70%.