EM-net: Deep learning for electron microscopy image segmentation

Abstract

Recent high-throughput electron microscopy techniques such as focused ion-beam scanning electron microscopy (FIB-SEM) provide thousands of serial sections which assist the biologists in studying sub-cellular structures at high resolution and large volume. Low contrast of such images hinder image segmentation and 3D visualisation of these datasets. With recent advances in computer vision and deep learning, such datasets can be segmented and reconstructed in 3D with greater ease and speed than with previous approaches. However, these methods still rely on thousands of ground-truth samples for training and electron microscopy datasets require significant amounts of time for carefully curated manual annotations. We address these bottlenecks with EMnet, a scalable deep convolutional neural network for EM image segmentation. We have evaluated EM-net using two datasets, one of which belongs to an ongoing competition on EM stack segmentation since 2012. We show that EM-net variants achieve better performances than current deep learning methods using small-and medium-sized ground-truth datasets. We also show that the ensemble of top EM-net base classifiers outperforms other methods across a wide variety of evaluation metrics.

EM-stellar: benchmarking deep learning for electron microscopy image segmentation

“…We performed an extensive survey of the literature to identify state-of-the-art deep neural networks that have been utilised for EM image segmentation. We chose CDeep3M [3], EMnet [4], PReLU-net [14], ResNet-50 [15], SegNet [16], U-net [17] and VGG-16 [18] for our experiments. Among these methods, we have experimented EM-net with all of its seven base classifiers bringing the total number of networks and methods to a maximum of thirteen.…”

“…Deep learning is a powerful approach to image segmentation that is being widely explored as a way to segment high-throughput biological datasets, including electron microscopy (EM) images [2]. In recent years, there have been several efforts to streamline the usage of such technologies for the community [3][4][5][6]. One crucial question that arises is whether we can use such platforms to segment all types of electron microscopy data and whether they have inherent limitations in segmenting particular types of ultrastructures.…”

“…We extracted 24 random patches from this dataset each having 512 × 512 pixels for training, testing and validation. After we manually annotated mitochondria on this sample, we split data randomly into training, validation and testing by 16 24 � , 4…”

“…All the experiments in this study except CDeep3M were implemented using TensorFlow GPU 1.8.0 CUDA 9.0 [29] and KERAS 2.2.4 [30]. These experiments were performed on a GPU cluster, HPC Spartan [31] as described in [4]. A stack of 100GB GPU instance was launched on AWS p3.2xlarge in the US West (Oregon) region to train CDeep3M.…”

“…� , and 4 24 � , respectively. The second dataset involves mice lateral habenula SBEM 1024 × 1024 × 80 voxels, as described previously [3].…”