Scholarlink

<p>3D semantic brain tumor segmentation is a fundamental task in medical imaging and healthcare. It aims to automate and facilitate the labeling of different brain structures and regions in three-dimensional medical images. Various deep learning models have demonstrated impressive segmentation capabilities on 3D biomedical image data sets. To harness the strengths of 3D models and potentially enhance segmentation accuracy, a novel ensemble learning approach rooted in deep learning has been introduced. Firstly, the proposal starts with a data pre-processing step including cropping, denoising, and data normalization. Secondly, the method employs a stacking technique based on five proposed semantic segmentation networks 3DU-Net, V-Net, 3DPSPNet, 3DW-Net, and 3DeepLabv3+ as training basic models. Finally, in the test phase, the obtained predictions will be stacked and fed as input to a meta-model for voxel-by-voxel multi-classification to achieve tumor segmentation. The meta-models used in most stacking ensemble learning methods are artificial neural networks and random forests to minimize computational performance and enhance accuracy. In our case, we introduce an optimized high meta-learner called 3DIncepTR considering these two purposes: computational performance and enhanced accuracy. Indeed, 3DIncepTR is an encoder–decoder network that employs a 3D-inception module and transformers for more deep, optimal, and accurate learning reached by point-wise convolutions with Inception and the self-attention mechanism in Transformers. The final model exhibits encouraging outcomes regarding the dice similarity coefficient (DSC) and other segmentation metrics. The evaluation was carried out on the available BraTS 2020 and 2018 datasets, where, in the first case, the approach attains DSC scores of 92.5%, 92%, and 87% for whole tumor (WT), core tumor (CT), and enhancing tumor (ET), respectively. In the second scenario, employing transfer learning from pre-trained models on the BraTS 2020 dataset yields success rates of 94.9%, 94.3%, and 89.7% for WT, CT, and ET, respectively.</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

3D-SELTML: stacking ensemble learning based on transformer meta-learner for 3D brain tumor segmentation

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Ilyasse Aboussaleh,
  • Jamal Riffi,
  • Khalid El Fazazy,
  • Adnane Mohamed Mahraz,
  • Hamid Tairi

摘要

3D semantic brain tumor segmentation is a fundamental task in medical imaging and healthcare. It aims to automate and facilitate the labeling of different brain structures and regions in three-dimensional medical images. Various deep learning models have demonstrated impressive segmentation capabilities on 3D biomedical image data sets. To harness the strengths of 3D models and potentially enhance segmentation accuracy, a novel ensemble learning approach rooted in deep learning has been introduced. Firstly, the proposal starts with a data pre-processing step including cropping, denoising, and data normalization. Secondly, the method employs a stacking technique based on five proposed semantic segmentation networks 3DU-Net, V-Net, 3DPSPNet, 3DW-Net, and 3DeepLabv3+ as training basic models. Finally, in the test phase, the obtained predictions will be stacked and fed as input to a meta-model for voxel-by-voxel multi-classification to achieve tumor segmentation. The meta-models used in most stacking ensemble learning methods are artificial neural networks and random forests to minimize computational performance and enhance accuracy. In our case, we introduce an optimized high meta-learner called 3DIncepTR considering these two purposes: computational performance and enhanced accuracy. Indeed, 3DIncepTR is an encoder–decoder network that employs a 3D-inception module and transformers for more deep, optimal, and accurate learning reached by point-wise convolutions with Inception and the self-attention mechanism in Transformers. The final model exhibits encouraging outcomes regarding the dice similarity coefficient (DSC) and other segmentation metrics. The evaluation was carried out on the available BraTS 2020 and 2018 datasets, where, in the first case, the approach attains DSC scores of 92.5%, 92%, and 87% for whole tumor (WT), core tumor (CT), and enhancing tumor (ET), respectively. In the second scenario, employing transfer learning from pre-trained models on the BraTS 2020 dataset yields success rates of 94.9%, 94.3%, and 89.7% for WT, CT, and ET, respectively.