Classification of neuronal morphology based on feature reconstruction and self-cure residual networks

Aiming at the problem of high morphological similarity between the different types of neurons and the large intra-class difference, which is easy to lead to low accuracy of neuron classification, a neural morphology classification method based on feature reconstruction and self-cure residual network is proposed. Firstly, to address the problems of edge pixel weakening and feature erosion by padding strategies that tend to occur during the convolution process of conventional convolution, a feature reconstruction module is constructed at the back end of the backbone network to retain important central features and filter damaged edge features. Then, the attention to neuronal morphological features is enhanced by using a self-attentive weight module and a rank regularization loss method, where the self-attention weight module assigns a weight to each sample to capture the sample importance for weighted loss. In addition, the rank regularization module re-ranked these weights in descending order, dividing them into two groups of high and low weights and regularizing the two groups by enforcing margins between the two average weights. The method achieved superior classification results on the NeuroMorpho-rat dataset, with twelve-way classification accuracies of 96.7%, 86.94% and 85.84% on the Img_raw, Img_resample and Img_XYalign datasets, separately. Comparing with the other methods, the present method has a higher classification accuracy of neurons. Comparing with the original ResNet18 network, it can effectively improve the neuron classification accuracy. 针对不同类别神经元之间的形态相似度高、类内区别性大, 容易导致神经元分类准确率不高的问题, 提出了一种基于特征重构自愈残差网络的神经元形态分类方法。针对传统卷积造成边缘像素弱化和填充策略带来新像素侵蚀特征的问题, 在基础网络后端构建特征重构模块来保留重要的中心特征并过滤受损的边缘特征。利用自注意力权重模块和排序正则化损失方法增强对神经元形态特征的关注。自注意力权重模块为每个样本重新分配权重, 以此捕获样本重要性进行加权损失; 排序正则化模块则将这些权重按降序重新排序, 分为高低2组权重, 同时通过在2组平均权重之间强制执行边距进行正则化处理。所提方法在大鼠神经元形态数据集上进行实验, 实现了较为优良的分类效果, 在Img_raw、Img_resample和Img_XYalign数据集上进行十二分类的准确率分别达到了96.7%, 86.94%, 85.84%。与其他分类方法相比, 所提方法具有更高的神经元形态分类准确率, 相较于基础网络ResNet18, 有效地提升了神经元形态分类准确率。