Scholarlink

<p>To enhance the quality of care for patients with brain tumors through improved pre-operative evaluation and post-operative follow-up monitoring, We have introduced an innovative hybrid model that leverages the successes of transformers in medical image segmentation. This model combines Vision Transformers with a 3D convolutional neural network (CNN) that utilizes attention mechanisms for the precise volumetric quantification of brain tumors in both pre-operative and post-operative scenarios. The model is named as Dual Attention Transformers (DATR). This architecture adopts transformers as the encoder to acquire sequence representations of the input volume, enabling it to capture long-range dependencies effectively. The transformer encoder is directly linked to the CNN decoder through skip connections at various resolutions to calculate the ultimate semantic segmentation output. The encoder and decoder are bridged by Dual Attention (DA) blocks at every network stage, which consists of the channel and spatial attention blocks for effectively emphasizing the relevant features and ignoring the irrelevant features. The network was trained and validated on two datasets: BraTS-2020 and BraTS-2021. Additionally, its performance was assessed on a private post-operative glioblastoma dataset from Uppsala University Hospital. The proposed network outperforms many competing networks for BraTS and the Uppsala dataset; the mean Dice scores for enhanced tumor (ET), whole tumor (WT) and tumor core (TC) obtained with the proposed network on BraTS 2020 and 2021 validation dataset are 0.7973, 0.8992, 0.8427 and 0.8338, 0.9110, 0.8680. On Uppsala University dataset the achieved Dice score for the post-operative residual tumor by DATR is 0.6712. The proposed DATR outperforms the state-of-the-art methods for the segmentation of pre- and post-operative brain tumors. The integration of transformer layers with DA enables the model to effectively capture shape-aware local details.</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

Transformers with Dual Attention for Volumetric Quantification of Pre and Post-operative Brain Tumors with Follow-Up

  • 期刊论文

  • 发表日期：

    2025年11月8日

查看全文
  • Swagata Kundu,
  • Dimitrios Toumpanakis,
  • Johan Wikstrom,
  • Robin Strand,
  • Ashis Kumar Dhara

摘要

To enhance the quality of care for patients with brain tumors through improved pre-operative evaluation and post-operative follow-up monitoring, We have introduced an innovative hybrid model that leverages the successes of transformers in medical image segmentation. This model combines Vision Transformers with a 3D convolutional neural network (CNN) that utilizes attention mechanisms for the precise volumetric quantification of brain tumors in both pre-operative and post-operative scenarios. The model is named as Dual Attention Transformers (DATR). This architecture adopts transformers as the encoder to acquire sequence representations of the input volume, enabling it to capture long-range dependencies effectively. The transformer encoder is directly linked to the CNN decoder through skip connections at various resolutions to calculate the ultimate semantic segmentation output. The encoder and decoder are bridged by Dual Attention (DA) blocks at every network stage, which consists of the channel and spatial attention blocks for effectively emphasizing the relevant features and ignoring the irrelevant features. The network was trained and validated on two datasets: BraTS-2020 and BraTS-2021. Additionally, its performance was assessed on a private post-operative glioblastoma dataset from Uppsala University Hospital. The proposed network outperforms many competing networks for BraTS and the Uppsala dataset; the mean Dice scores for enhanced tumor (ET), whole tumor (WT) and tumor core (TC) obtained with the proposed network on BraTS 2020 and 2021 validation dataset are 0.7973, 0.8992, 0.8427 and 0.8338, 0.9110, 0.8680. On Uppsala University dataset the achieved Dice score for the post-operative residual tumor by DATR is 0.6712. The proposed DATR outperforms the state-of-the-art methods for the segmentation of pre- and post-operative brain tumors. The integration of transformer layers with DA enables the model to effectively capture shape-aware local details.