FusionNet: A deep fully residual convolutional neural network for image segmentation in connectomics (1612.05360v2)

Abstract: Electron microscopic connectomics is an ambitious research direction with the goal of studying comprehensive brain connectivity maps by 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. Recently, deep learning has drawn much attention in computer vision because of its exceptional performance in image classification tasks. For this reason, its application to connectomic analyses holds great promise, as well. In this paper, we introduce a novel deep neural network architecture, FusionNet, for the automatic segmentation of neuronal structures in connectomics data. FusionNet leverages the latest advances in machine learning, such as semantic segmentation and residual neural networks, with the novel introduction of summation-based skip connections to allow a much deeper network architecture for a more accurate segmentation. We demonstrate the performance of the proposed method by comparing it with state-of-the-art electron microscopy (EM) segmentation methods from the ISBI EM segmentation challenge. We also show the segmentation results on two different tasks including cell membrane and cell body segmentation and a statistical analysis of cell morphology.

• The paper introduces FusionNet, a deep fully residual CNN with summation-based skip connections that mitigate vanishing gradients in deep architectures.
• The paper demonstrates superior segmentation performance in the ISBI 2012 EM challenge, outperforming state-of-the-art methods like U-net with reduced post-processing.
• The paper employs a tailored data augmentation strategy using orientation variants and elastic deformations to improve model robustness on EM datasets.

Overview of FusionNet: A Deep Fully Residual Convolutional Neural Network for Image Segmentation in Connectomics

The paper presents FusionNet, a novel deep neural network architecture designed to address the demanding requirements of automated neuronal segmentation in connectomics, the paper of brain connectivity maps. This work leverages recent advancements in deep learning to enhance the segmentation accuracy in electron microscopic (EM) connectomics datasets. A crucial innovation is the introduction of summation-based skip connections within a residual fully convolutional network (FCN), enabling the construction of deeper networks with improved segmentation performance.

Key Contributions

1. Network Architecture: FusionNet is an extension of the U-net and residual CNN frameworks. By incorporating summation-based skip connections, it effectively mitigates issues related to vanishing gradients, thus facilitating the training of deeper networks. These skip connections are pivotal in maintaining information flow across multiple layers and enhancing the capacity of the network to capture complex features in EM data.
2. Comparative Performance Evaluation: The authors validate FusionNet through comparative analyses against state-of-the-art EM segmentation methods as demonstrated by its superior results in the ISBI 2012 EM segmentation challenge. FusionNet's architecture outperformed significant existing methodologies, such as U-net and techniques employing extensive post-processing, by offering competitive segmentation accuracy with minimal post-processing.
3. Data Enrichment Techniques: FusionNet introduces a data augmentation strategy tailored for EM data, which involves generating orientation variants to extend the training dataset effectively. This approach enhances the robustness of the segmentation by incorporating diverse image orientations and elastic field deformations to improve generalization.
4. Case Studies: The flexibility of the FusionNet framework is showcased through its application to distinct EM segmentation tasks, specifically targeting the segmentation of cell membranes and neuronal nuclei in larval zebrafish. FusionNet demonstrates reduced false-positive rates and increases in accuracy over competing methods.

Implications and Future Directions

The deployment of FusionNet in the analysis of large-scale EM datasets like those in connectomics research portends significant efficiency gains. By facilitating more accurate segmentation with reduced human annotation effort, it aids the scaling of connectomic studies to larger datasets, helping to unravel the complex neuronal architectures in the brain. The residual learning paradigm adopted here could influence further development in designing deep architectures for similar high-fidelity image analysis tasks.

In conclusion, while FusionNet constitutes a meaningful step forward in connectomics image segmentation, the potential for further exploration remains. Incorporating distributed training strategies or exploring even deeper network architectures could yield insights into extending FusionNet's capabilities. Additionally, testing and validating cross-domain applications, such as in medical imaging modalities, could further enhance its versatility and impact on real-world image analysis tasks.