Automated Neuron Tracing with Reinforcement Learning and Branch Point Classification
Poster No:
1549
Submission Type:
Abstract Submission
Authors:
Bryson Gray1, Daniel Tward1
Institutions:
1University of California, Los Angeles, Los Angeles, CA
First Author:
Co-Author:
Introduction:
Quantifying sincle-cell morphology is crucial to understanding brain circuit formation, but dataset size and neuron complexity make automated processing essential. While various automatic neuron tracing methods exist, the tracing of dense neurite trees remains a challenge. Recently, Dai et al. (2019) introduced a deep reinforcement learning (RL) method for tracing neurons. RL, in applicable settings, outmatches supervised learning in needing no labeled data, instead exploring an environment and learning from feedback given by a reward function. RL offers potential in robustness to noise and the ability to build prior knowledge into the reward function. However, existing implementations are restricted to 2D images, and lack branching due to the sequential nature of RL agents (Dai et al., 2019), (Balaram et al., 2019). Here we present our novel RL method capable of tracing branching neurons in 3D.
Methods:
We built a neuron tracing module composed of two separately trained deep convolutional neural networks. The first is a policy network, which inputs a 3D image region and the currently traced path, and outputs the parameters of a multivariate Gaussian, from which the direction of a next step is sampled. The path terminates when the next step chosen is near 180 degrees from the previous step. The second component is a branch classifier, which marks the locations of branches as they are encountered by the tracking agent, and adds them to a queue. Once the agent finishes one path, it begins tracing from the next branch point if any are left in the queue. We created synthetic data based on 73 real neuron morphology trees randomly selected and downloaded from the publicly available dataset at neuromorpho.org. The vectorized data were translated into synthetic 3D neuron images by drawing line segments between every node point and blurring the line with a Gaussian filter. Each image was stacked into three RGB channels with independent random noise added to each channel. The tracing seed point for each neuron was set to the first point in the tree. Branch points were also extracted and used to make branch masks, binary images where a small region surrounding each branch point was labeled positive. The policy network was optimized using the Soft Actor Critic algorithm (Haarnoja et al., 2018), which aims to maximize the discounted sum of rewards over episodes. The reward is a function of the change in sum of square error between the estimated path image and the true neuron image plus a term to enforce smoothness. We used a 34-layer ResNet for the branch classifier. The training dataset was created by randomly sampling image windows around the neurons and assigning target labels based on the branch mask. We included random permutations and flips to augment the input data to minimize overfitting and utilized class balancing for generalizability. For evaluation we computed the percent of neuron traced.
Results:
Reconstruction accuracies are highly variable, so here we report the results of neuron tracing on two representative neurons. Figure 1A shows an example neuron which was completely traced (97.9% coverage). However, errors in branch identification sometimes led to large sections being untraced as in Figure 1B with 60.3% coverage. Performing the same reconstructions without branching leads to dramatically lower coverage by comparison (22.7% and 3.2% coverage respectively) (Fig. 1C,D).
Conclusions:
Our results demonstrate the feasibility of 3D neuron tracing with RL. Comparison between tracing with and without the branch classifier reveals the importance of incorporating branching in the model. The difficulty of tracing depends on the density of neurites and of their split-ups, as the branch classifier currently classifies bifurcations but not multiple branches. Future work should focus on improving branch classification accuracy, dealing with multiple branches at a point, and generalizing the model to perform on real microscopy images, rather than synthetic data.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 2
Image Registration and Computational Anatomy
Methods Development 1
Segmentation and Parcellation
Keywords:
Cellular
Computational Neuroscience
Machine Learning
Modeling
Morphometrics
Neuron
Tractography
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
Provide references using APA citation style.
Dai, T. et al., (2019). Deep Reinforcement Learning for Subpixel Neural Tracking. Proceedings of Machine Learning Research, 102, 130-150.
Haarnoja, T. et al., (2018). Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. International Conference on Machine Learning, abs/1801.01290, 1861-1870.