神经元图像大数据的弱监督自动识别技术

Weakly supervised automatic recognition technology for neuron image big data

在介观尺度上,小鼠大脑图像的数据量可达到10 TB量级,人脑数据量则达到惊人的几十PB,从海量脑图像数据中识别和分析神经元的形态是一项复杂且具有挑战的任务。当前研究人员提出了基于传统机器学习和深度学习的神经元识别算法,其中传统机器学习方法存在迁移、泛化能力较差的问题,基于深度学习的算法虽然可以通过海量精确标注的训练数据提高模型的泛化性,但缺乏精确且丰富的图像标记数据集,因此同样存在过拟合和泛化能力弱等问题。本文提出了一种基于深度学习的弱监督神经元识别方案,仅需要少量有标注的数据,即可通过迭代策略获取海量神经元图像的精确识别结果,具备较强的泛化能力,并最大限度减少人工参与量。该方法在fMOST、BigNeuron等数据集上进行了实验,自动识别精度F1值分别为0.9247和0.8318,优于其他对比的神经元识别算法。

At the mesoscale level,the data volume of mouse brain images can reach the level of 10 TB,while that of human brain can reach the astonishing level of several dozen PB.Recognition and analyzing the morphology of neurons from massive brain image data is a complex and challenging task.Currently,researchers have proposed neuron recognition algorithms based on both traditional machine learning and deep learning methods.Traditional machine learning methods suffer from poor transfer and generalization ability,while deep learning-based algorithms,although able to improve the model’s generalization ability through massive accurately labeled training data,lack precise and rich image annotation datasets,resulting in problems such as overfitting and weak generalization ability.This article proposes a weakly supervised neuron recognition method based on deep learning,which requires only a small amount of labeled data,and can obtain accurate recognition results of a large amount of unlabeled data through iterative strategies,which has strong generalization capabilities and minimizes human participation.This method has been experimentally verified on brain image big data such as fMOST and BigNeuron,with recognition accuracies of F1 scores of 0.9247 and 0.8318,respectively,which are better than other compared neuron recognition algorithms.