### Introduction

Reconstruction of 3D neuron morphology is essential in brain studies. In the nervous system, electrical impulses and chemical materials are transported through pathways between connected neurons. Characterizing the 3D morphology of individual neurons is fundamentally important for understanding neuronal identity, morphological analysis, determining potential connectivity and promoting the study of brain-related diseases like Alzheimer. Despite a number of existing studies, this task still remains challenging, especially when the 3D microscopic images have low single-to-noise ratio and noncontinuous segments of neurite patterns. We are seeking a more powerful method for neuron population reconstruction on our own collected mouse brain images

### Dataset

Though many neuron tracing techniques have been proposed, no dataset of large-scale dense neuronal population has been published up to now. In order to validate our approach, and most importantly, to support further studies on the dense neuronal population reconstruction, we construct a dataset of 3D optical microscopy images.

Our VISoR-40 dataset consists of 40 volumetric images captured from a mouse brain. The whole-brain images were captured by the VISoR imaging system, at the physical resolution of 0.5×0.5×0.5 um^3 per voxel, and image resolution of about 20000× 30000×24000. At this scale, identiﬁcation of every individual neuron is feasible for neuron morphology analysis. The raw volumetric data has 16-bit dynamic range of intensity that preserves enough neurite details.

For training the segmentation network, we randomly pick 4/5 of the entire dataset as training dataset. Then the remaining eight subjects were used as the test set and validation set to evaluate our method with diﬀerent settings. To get manual annotations of the eight subjects, we ask two expert annotators to manually reconstruct individual neurons using the 3D Virtual Finger plugin in Vaa3D. We can use the Vaa3D to see the volumetric images and load the SWC ﬁles to see the manual annotations and reconstruction results.

### Evaluation Metrics

To quantitatively evaluate the performance of neuron population reconstruction, four commonly used metrics, including Precision, Recall, F-Score, and Jaccard, are computed to measure the diﬀerence between a reconstruction and the ground truth. Their deﬁnitions are deﬁned as follows:

$$\mathit{Precission(R,G)=\frac {\vert R \bigcap G \vert}{\vert R \vert}}$$

$$\mathit{Recall(R,G)=\frac {\vert R \bigcap G \vert}{\vert G \vert}}$$

$$\mathit{F-Score(R,G)=\frac {2\vert R \bigcap G \vert}{\vert R \vert + \vert G \vert}}$$

$$\mathit{Jaccard(R,G)=\frac {\vert R \bigcap G \vert}{\vert R \bigcup G \vert}}$$

where R denotes the set of points on the reconstructed neurons, G denotes the set of neuron points in the ground truth, |·| denotes the number of points in a set. The four metrics are first computed on each test image according to the manually labeled skeleton, then averaged in test images weighted by the total length of the neuronal processes in each image.

Citation: If you find our dataset or code is useful for your research, please cite: J. Zhao, X. Chen, Z. Xiong, D. Liu, J. Zeng, Y. Zhang, Zheng-Jun Zha, G. Bi and F. Wu, “Progressive learning for neuronal population reconstruction from optical microscopy images,” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2019: 750–759.

jzhaoch@mail.ustc.edu.cn