基于Focal Recall Loss的阿尔兹海默症病灶分割模型

The Alzheimer’s Disease Segmentation Model Based on Focal Recall Loss

摘要: 目前的阿尔兹海默症病灶分割算法主要以深度分割网络作为主流算法对组织病理区域进行分割，但是这些方法在面临类别不均衡的情况时，整体性能会受少数类的影响而陷入局部最优的情况。针对上述情况，首先，本文提出一种新的语义分割模型，名为Global Attention UNet (GAUNet)，该模型嵌入了全局注意力模块以及组线性层模块对特征空间中的通道信息以及空间信息进行挖掘，从而提高模型的特征表示能力。其次，针对类别不均衡的问题，本文提出一种局部回归损失函数(Focal Recall Loss)，针对每轮次召回情况动态调整各个类别的权重，从而使得模型更关注少数类的类别特征信息。本文所提出的方法在Alzheimer’s Disease Neuroimaging Initiative (ADNI)数据集中对6种组织类别区域(额叶、颞叶、顶叶、海马体、中脑、半卵圆中心)同时进行分割，与当下的模型相比，本文提出的方法在少数类别半卵圆中心的IOU比当前最新方法高出6.19%。

Abstract: Currently, deep segmentation approaches become the main algorithms for lesion segmentation in Alzheimer’s Disease. However, these methods hardly extract the edge information of regions that belonged to the minority category. In order to solve the above problem, firstly, we propose a novel segmentation model, called Global Attention UNet (GAUNet). Here, we embedded a Global Attention Block into our proposed model to explore the channel and spatial feature to improve the feature discriminative. Secondly, we propose the Focal Recall Loss function to alleviate the problem of the imbalance category. Specifically, our loss function can dynamically adjust the weights of each category based on each epoch of recall ratios. Finally, our proposed method applies to the Alz-heimer’s Disease Neuroimaging Initiative (ADNI) dataset to segment six regions including the Frontal Lobe, Temporal Lobe, Parietal Lobe, Hippocampus, Midbrain and Semi-oval center. The ex-perimental results show that our proposed method achieves 6.19% higher IOU in the category of semi-oval center than several state-of-the-arts.