Cross-Classification Clustering: An Efficient Multi-Object Tracking Technique for 3-D Instance Segmentation in Connectomics

Meirovitch, Yaron; Lu, Mi; Saribekyan, Hayk; Matveev, Alexander; Rolnick, David; Shavit, Nir

doi:10.1109/cvpr.2019.00862

Cited by 38 publications

(51 citation statements)

References 56 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In brief, the algorithm expanded all skeleton nodes in each section until they fully filled the images of all labelled cells. Cellular borders were predicted by a shallow Convolutional Neural Network (CNN) that builds on XNN 55,56 , a recently developed high performance system which computes convolutions on CPUs, to achieve border prediction throughput of 10MB/s 57,58 . Node expansion was computed with a dedicated Cilk-based code 59 that parallelized the Dijkstra graph search algorithm.…”

Section: Segmentation For Volumetric Reconstructionmentioning

confidence: 99%

Connectomes across development reveal principles of brain maturation

Witvliet

Mulcahy

Mitchell

et al. 2020

Preprint

Self Cite

160

View full text Add to dashboard Cite

From birth to adulthood, an animal's nervous system changes as its body grows and its behaviours mature. However, the extent of circuit remodelling across the connectome is poorly understood. Here, we used serial-section electron microscopy to reconstruct the brain of eight isogenic C. elegans individuals at different ages to learn how an entire wiring diagram changes with maturation. We found that the overall geometry of the nervous system is preserved from birth to adulthood, establishing a constant scaffold upon which synaptic change is built. We observed substantial connectivity differences among individuals that make each brain partly unique. We also observed developmental connectivity changes that are consistent between animals but different among neurons, altering the strengths of existing connections and creating additional connections. Collective synaptic changes alter information processing of the brain. Across maturation, the decision-making circuitry is maintained whereas sensory and motor pathways are substantially remodelled, and the brain becomes progressively more modular and feedforward. These synaptic changes reveal principles that underlie brain maturation.Witvliet et al. | bioRχiv | May 21, 2020 | 1-26

show abstract

Section: Segmentation For Volumetric Reconstructionmentioning

confidence: 99%

Connectomes across development reveal principles of brain maturation

Witvliet

Mulcahy

Mitchell

et al. 2020

Preprint

Self Cite

160

View full text Add to dashboard Cite

show abstract

“…State-of-the-art automated segmentation techniques use deep convolutional neural networks (DCNNs) to trace ultrastructures in 3D-EM images 7–11 . A DCNN-based segmentation technique generally comprises two steps: semantic- and instance segmentation.…”

Section: Introductionmentioning

confidence: 99%

“…10 suggested flood-filling networks (FFNs), a single-object tracking technique, and Meirovitch et al . 11 introduced cross-classification clustering, a multi-object tracking technique, merging the semantic- and instance segmentation in recurrent neural networks. These techniques have a bottom-up design, where recurrent networks maintain the prediction for the object shape and learn to reconstruct neuronal processes with more plausible shapes.…”

Section: Introductionmentioning

confidence: 99%

“…Distinctive image features are required for a segmentation technique with a bottom-up design that is subjected to greedy optimization, making the locally optimal choice at each stage while searching for a global optimum. Therefore, the mentioned automated techniques 7–11 cannot be used to segment low-resolution images. For example, techniques such as DeepEM3D 8 and its cloud-based implementation 7 essentially rely on a precise semantic segmentation and apply no mechanism to correct potential topological errors during instance segmentation.…”

Section: Introductionmentioning

confidence: 99%

“…Therefore, semantic segmentation errors propagate into instance segmentation as either over- or under-segmentation. Techniques such as FFN 10 and its multi-object tracking counterpart 11 , where networks learn the shape of a neural process, would experience over-segmentation errors. Merging FFN super-voxels does not necessarily generate a correct segmentation of an axon as the segmentation leaks to the extra-axonal space at nodes of Ranvier.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

DeepACSON: Automated Segmentation of White Matter in 3D Electron Microscopy

Abdollahzadeh

Belevich

Jokitalo

et al. 2019

Preprint

View full text Add to dashboard Cite

Advances in 3-dimensional electron microscopy (3D-EM) have made possible imaging brain ultrastructures with nanoscopic resolutions from hundreds of micrometers tissue. To quantify the acquired big datasets, it requires automated segmentation techniques that annotate the entirety of ultrastructural components. Recently, deep convolutional neural networks (DCNNs) have emerged as the key to automated human-level segmentation of neurites in 3D-EM datasets. Current techniques, however, address the segmentation of high-resolution datasets and for small numbers of neurites. Therefore, we present DeepACSON.DeepACSON performs DCNN-based semantic segmentation and 3D shape analysis based instance segmentation. We used DeepACSON to segment ten low-resolution serial block-face scanning electron microscopy datasets of the white matter of rats after sham-operation and traumatic brain injury. DeepACSON segmented hundreds of thousands of long-span myelinated axons, thousands of cell nuclei, and millions of mitochondria with excellent evaluation scores. DeepACSON quantified the morphological aspects of myelinated axons, such as axonal diameter, eccentricity, and tortuosity, as well as the spatial organization of ultrastructures, such as inter-mitochondrial distance, and the density of cells and myelinated axons.15 × 15 × 15 µm 3 white matter ultrastructures requires 2 400 h of an expert's time 4 . Semi-automated segmentation methods based on machine learning approaches 3, 5 have improved the rate of segmentation. However, these methods still require a considerable amount of manual interaction as the segmentation is driven on the manually extracted skeletons of neurites, proofreading, and correction of errors.1 Deep convolutional neural networks (DCNNs) have emerged as the key to the state-of-the-art automated segmentation techniques 6-9 . DCNN-based segmentation methods generally comprise two steps: generating a membrane probability map from 3D-EM datasets referred to as semantic segmentation, followed by instance segmentation that turns probability maps into individual object instances. DCNNs are typically used for semantic segmentation and more traditional image analysis techniques favoring bottom-up design, i.e., over-segmentation and subsequent merge, are used for instance segmentation. For example, DeepEM3D 7 applied 3D watershed transform on the membrane probability maps. Funke et al. 8 proposed learning voxel affinities, followed by 3D watershed and iterative region agglomeration on the watershed segmentation. Only recently, Januszewski et al. 9 suggested flood-filling networks merging the semantic and instance segmentation into a recurrent neural network. Flood-filling networks have a bottom-up design, offering a sophisticated training-based alternative for over-segmentation and subsequent merge.Bottom-up design of the state-of-the-art EM segmentation techniques 6-9 is subjected to local optimization. Hence, segmentation requires distinctive image features such as high contrast between cellular membranes and other struc...

show abstract