BraTS-2019

自动脑肿瘤分割方法在脑肿瘤诊断和治疗的全过程中起着极其重要的作用。在本文中，我们提出了一种多步级联网络，该网络考虑了脑肿瘤子结构的分层拓扑，并将子结构从粗到细进行了细分。在分割过程中，前一步骤的结果被用作下一步的先验信息，以指导更精细的分割过程。整个网络都以端到端的方式进行培训。此外，为了缓解梯度消失的问题并减少过拟合，我们添加了几个辅助输出作为每个步骤的深度监督，并分别引入了几种数据增强策略，这对于脑肿瘤分割是非常有效的。最后，利用局灶性损失来解决肿瘤区域和背景明显失衡的问题。我们的模型是在BraTS 2019验证数据集上进行测试的，平均dice系数的初步结果分别是0.886，0.813，0.771对整个肿瘤，肿瘤核心和增强肿瘤。

Automatic brain tumor segmentation methods play an extremely critical role in the entire process of brain tumor diagnosis and treatment. In this paper, we propose a multi-step cascaded network that considers the hierarchical topology of brain tumor substructures and subdivides these substructures from coarse to fine. During the segmentation procedure, the outputs of the previous step are employed as prior information for the subsequent step to guide more precise segmentation. The entire network is trained in an end-to-end manner. Furthermore, to mitigate the gradient vanishing issue and alleviate overfitting, we introduce several auxiliary outputs as deep supervision for each step, and separately adopt multiple data augmentation strategies that prove highly effective for brain tumor segmentation. Finally, Focal Loss is utilized to address the pronounced class imbalance between tumor regions and the background. Our model is evaluated on the BraTS 2019 validation dataset, and the preliminary mean Dice coefficient results for the whole tumor, tumor core, and enhancing tumor are 0.886, 0.813, and 0.771, respectively.

搜集汇总

背景与挑战

背景概述

BraTS-2019是一个专注于自动脑肿瘤分割的数据集，由哈尔滨工业大学于2019年发布。该数据集采用多步级联网络方法，通过端到端训练和深度监督技术，有效处理脑肿瘤子结构的分割，并利用数据增强和局灶性损失函数来优化模型性能。在验证测试中，模型对整体肿瘤、肿瘤核心和增强肿瘤的分割dice系数分别达到0.886、0.813和0.771，显示出较高的分割准确性。