# FusionNet: A deep fully residual convolutional neural network for image segmentation in connectomics

**Authors:** Tran Minh Quan, David G. C. Hildebrand, Won-Ki Jeong  
**arXiv:** [1612.05360](https://arxiv.org/abs/1612.05360)  
**Date:** 2016-12-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1612.05360v2/

## Key points

- FusionNet is a deep neural network architecture for automatic segmentation of neuronal structures in connectomics data.
- Connectomics aims to study brain connectivity using high-throughput, nano-scale microscopy.
- Deep learning has shown exceptional performance in image classification tasks and holds promise for connectomic analyses.
- FusionNet incorporates semantic segmentation, residual neural networks, and summation-based skip connections for improved accuracy.
- The paper compares FusionNet with state-of-the-art electron microscopy (EM) segmentation methods from the ISBI EM segmentation challenge.
- FusionNet is demonstrated on cell membrane and cell body segmentation tasks, with statistical analysis of cell morphology.
- The authors used FusionNet to analyze 16,000 sections of the larval zebrafish brain and identified approximately 180,000 cell bodies.
- Visualizations of cross-sectional views of the EM volume overlaid with cell body segmentations are included.
- FusionNet is presented as an effective approach for automatic neuronal structure segmentation in connectomics data, surpassing existing methods.
- The findings contribute to advancing our understanding of brain connectivity and have potential applications in neuroscience research and medical diagnostics.

## Summary

The paper titled "FusionNet: A deep fully residual convolutional neural network for image segmentation in connectomics" introduces a novel deep neural network architecture called FusionNet for the automatic segmentation of neuronal structures in connectomics data. Connectomics is a research direction that aims to study comprehensive brain connectivity maps using high-throughput, nano-scale microscopy. One of the main challenges in connectomics research is developing scalable image analysis algorithms that require minimal user intervention. Deep learning has gained significant attention in computer vision due to its exceptional performance in image classification tasks. Therefore, applying deep learning to connectomic analyses holds great promise. FusionNet leverages the latest advances in machine learning, such as semantic segmentation and residual neural networks, with the introduction of summation-based skip connections. These skip connections allow for a much deeper network architecture, leading to more accurate segmentation results. The proposed method's performance is demonstrated by comparing it with state-of-the-art electron microscopy (EM) segmentation methods from the ISBI EM segmentation challenge. The authors also showcase the segmentation results on two different tasks: cell membrane and cell body segmentation. Additionally, they perform a statistical analysis of cell morphology. In their experiments, the authors deployed the trained network to analyze a complete set of 16,000 sections of the larval zebrafish brain imaged at high resolution. They discovered approximately 180,000 cell bodies in the larval zebrafish brain using FusionNet and included visualizations of cross-sectional views of the EM volume overlaid with cell body segmentations. Overall, this paper presents FusionNet as an effective deep learning approach for automatic neuronal structure segmentation in connectomics data which demonstrates its superiority over existing methods and provides insights into cell morphology through statistical analysis. The findings contribute to advancing our understanding of brain connectivity and have potential applications in neuroscience research and medical diagnostics.