This repository contains the implementation of my Master thesis at EPFL described in the project report. The implementation code is private and has not been publicly released so all the scripts are left empty on this repository.
Reconstruction of neuron morphologies has been a longstanding challenge for the computer vision community. The broad spectrum of its applications, ranging from providing more accurate neuron simulation models to neurological disorders identification, highlights the multidisciplinary aspect of this research topic and its significant potential to help science move forward. Considerable effort has been put towards developing tools that can reduce the excessive amount of manual input that has to be provided to complete a reconstruction. However, current methods have struggled to revolutionize the field, mostly failing to keep the tree topology of the reconstructions nor providing the output in the form of a graph, preferred for many uses. To this end, we propose a new approach to the neuron tracing problem that can be generalized to other applications such as vasculature and road network extraction. The designed method outputs a graph-based reconstruction, which provides branch-specific information, extracts the bifurcation points and preserves the tree structure of the reconstructed neuron. We design a model that can easily be integrated to an annotation pipeline, in which a Reinforcement Learning agent is trained to navigate through neuron images and trace the underlying branch trajectories. We show that the proposed approach can be adapted to work on synthetic data, first by applying in two dimensions, and then by demonstrating its applicability to neuronal structures in three dimensions. The approach is finally tested on predictions coming from real neuron microscopy scans. The agent is able to move along neurites on real data, but missing information and significant gaps in the annotations expose the need for more accurate annotations for the complete assessment of the applicability of the proposed approach to real data use cases.
In red: the true segmentation, and in green: the predicted reconstruction.