Search

Studies of individual neurons in the Drosophila nervous system are facilitated by transgenic lines that sparsely and repeatably label respective neurons of interest. Sparsity can be enhanced by means of intersectional approaches like the split-GAL4 system, which labels the positive intersection of the expression patterns of two (denser) GAL4 lines. To this end, two GAL4 lines have to be identified as labelling a neuron of interest. Current approaches to tackling this task include visual inspection, as well as automated search in 2d projection images, of single cell multi-color flip-out (MCFO) acquisitions of GAL4 expression patterns. There is to date no automated method available that performs full 3d search in MCFO imagery of GAL4 lines, nor one that leverages automated reconstructions of the labelled neuron morphologies. To close this gap, we propose PatchPerPixMatch, a fully automated approach for finding a given neuron morphology in MCFO acquisitions of Gen1 GAL4 lines. PatchPerPixMatch performs automated instance segmentation of MCFO acquisitions, and subsequently searches for a target neuron morphology by minimizing an objective that aims at covering the target with a set of well-fitting segmentation fragments. Patch-PerPixMatch is computationally efficient albeit being full 3d, while also highly robust to inaccuracies in the automated neuron instance segmentation. We are releasing PatchPerPixMatch search results for ~30,000 neuron morphologies from the Drosophila hemibrain in ~20,000 MCFO acquisitions of ~3,500 Gen1 GAL4 lines.Codehttps://github.com/Kainmueller-Lab/PatchPerPixMatchResultshttps://pppm.janelia.org

The neural circuits responsible for behavior remain largely unknown. Previous efforts have reconstructed the complete circuits of small animals, with hundreds of neurons, and selected circuits for larger animals. Here we (the FlyEM project at Janelia and collaborators at Google) summarize new methods and present the complete circuitry of a large fraction of the brain of a much more complex animal, the fruit fly Drosophila melanogaster. Improved methods include new procedures to prepare, image, align, segment, find synapses, and proofread such large data sets; new methods that define cell types based on connectivity in addition to morphology; and new methods to simplify access to a large and evolving data set. From the resulting data we derive a better definition of computational compartments and their connections; an exhaustive atlas of cell examples and types, many of them novel; detailed circuits for most of the central brain; and exploration of the statistics and structure of different brain compartments, and the brain as a whole. We make the data public, with a web site and resources specifically designed to make it easy to explore, for all levels of expertise from the expert to the merely curious. The public availability of these data, and the simplified means to access it, dramatically reduces the effort needed to answer typical circuit questions, such as the identity of upstream and downstream neural partners, the circuitry of brain regions, and to link the neurons defined by our analysis with genetic reagents that can be used to study their functions.Note: In the next few weeks, we will release a series of papers with more involved discussions. One paper will detail the hemibrain reconstruction with more extensive analysis and interpretation made possible by this dense connectome. Another paper will explore the central complex, a brain region involved in navigation, motor control, and sleep. A final paper will present insights from the mushroom body, a center of multimodal associative learning in the fly brain.

We present a novel method for proposal free instance segmentation that can handle sophisticated object shapes which span large parts of an image and form dense object clusters with crossovers. Our method is based on predicting dense local shape descriptors, which we assemble to form instances. All instances are assembled simultaneously in one go. To our knowledge, our method is the first non-iterative method that yields instances that are composed of learnt shape patches. We evaluate our method on a diverse range of data domains, where it defines the new state of the art on four benchmarks, namely the ISBI 2012 EM segmentation benchmark, the BBBC010 C. elegans dataset, and 2d as well as 3d fluorescence microscopy data of cell nuclei. We show furthermore that our method also applies to 3d light microscopy data of Drosophila neurons, which exhibit extreme cases of complex shape clusters.

Probabilistic convolutional neural networks, which predict distributions of predictions instead of point estimates, led to recent advances in many areas of computer vision, from image reconstruction to semantic segmentation. Besides state of the art benchmark results, these networks made it possible to quantify local uncertainties in the predictions. These were used in active learning frameworks to target the labeling efforts of specialist annotators or to assess the quality of a prediction in a safety-critical environment. However, for instance segmentation problems these methods are not frequently used so far. We seek to close this gap by proposing a generic method to obtain model-inherent uncertainty estimates within proposal-free instance segmentation models. Furthermore, we analyze the quality of the uncertainty estimates with a metric adapted from semantic segmentation. We evaluate our method on the BBBC010 C. elegans dataset, where it yields competitive performance while also predicting uncertainty estimates that carry information about object-level inaccuracies like false splits and false merges. We perform a simulation to show the potential use of such uncertainty estimates in guided proofreading.

Segmentation and tracking of epithelial cells in light microscopy (LM) movies of developing tissue is an abundant task in cell- and developmental biology. Epithelial cells are densely packed cells that form a honeycomb-like grid. This dense packing distinguishes membrane-stained epithelial cells from the types of objects recent cell tracking benchmarks have focused on, like cell nuclei and freely moving individual cells. While semi-automated tools for segmentation and tracking of epithelial cells are available to biologists, common tools rely on classical watershed based segmentation and engineered tracking heuristics, and entail a tedious phase of manual curation. However, a different kind of densely packed cell imagery has become a focus of recent computer vision research, namely electron microscopy (EM) images of neurons. In this work we explore the benefits of two recent neuron EM segmentation methods for epithelial cell tracking in light microscopy. In particular we adapt two different deep learning approaches for neuron segmentation, namely Flood Filling Networks and MALA, to epithelial cell tracking. We benchmark these on a dataset of eight movies with up to 200 frames. We compare to Moral Lineage Tracing, a combinatorial optimization approach that recently claimed state of the art results for epithelial cell tracking. Furthermore, we compare to Tissue Analyzer, an off-the-shelf tool used by Biologists that serves as our baseline.

We propose a novel label fusion technique as well as a crowdsourcing protocol to efficiently obtain accurate epithelial cell segmentations from non-expert crowd workers. Our label fusion technique simultaneously estimates the true segmentation, the performance levels of individual crowd workers, and an image segmentation model in the form of a pairwise Markov random field. We term our approach image-aware STAPLE (iaSTAPLE) since our image segmentation model seamlessly integrates into the well-known and widely used STAPLE approach. In an evaluation on a light microscopy dataset containing more than 5000 membrane labeled epithelial cells of a fly wing, we show that iaSTAPLE outperforms STAPLE in terms of segmentation accuracy as well as in terms of the accuracy of estimated crowd worker performance levels, and is able to correctly segment 99% of all cells when compared to expert segmentations. These results show that iaSTAPLE is a highly useful tool for crowd sourcing image segmentation., Accepted to ISBI2017

We consider the task of pixel-wise semantic segmentation given a small set of labeled training images. Among two of the most popular techniques to address this task are Decision Forests (DF) and Neural Networks (NN). In this work, we explore the relationship between two special forms of these techniques: stacked DFs (namely Auto-context) and deep Convolutional Neural Networks (ConvNet). Our main contribution is to show that Auto-context can be mapped to a deep ConvNet with novel architecture, and thereby trained end-to-end. This mapping can be used as an initialization of a deep ConvNet, enabling training even in the face of very limited amounts of training data. We also demonstrate an approximate mapping back from the refined ConvNet to a second stacked DF, with improved performance over the original. We experimentally verify that these mappings outperform stacked DFs for two different applications in computer vision and biology: Kinect-based body part labeling from depth images, and somite segmentation in microscopy images of developing zebrafish. Finally, we revisit the core mapping from a Decision Tree (DT) to a NN, and show that it is also possible to map a fuzzy DT, with sigmoidal split decisions, to a NN. This addresses multiple limitations of the previous mapping, and yields new insights into the popular Rectified Linear Unit (ReLU), and more recently proposed concatenated ReLU (CReLU), activation functions.

Accurate localization, identification and segmentation of vertebrae is an important task in medical and biological image analysis. The prevailing approach to solve such a task is to first generate pixelindependent features for each vertebra, e.g. via a random forest predictor, which are then fed into an MRF-based objective to infer the optimal MAP solution of a constellation model. We abandon this static, twostage approach and mix feature generation with model-based inference in a new, more flexible, way. We evaluate our method on two data sets with different objectives. The first is semantic segmentation of a 21-part body plan of zebrafish embryos in microscopy images, and the second is localization and identification of vertebrae in benchmark human CT.

In this chapter, we propose fully automatic segmentation pipelines for three segmentation problems that appear in clinical practice. Pipelines are assembled from methods described in Chapter 3 (Sections 3.1 – 3.4).

In Section 3.5 we described multi-object GraphCuts and proposed a mesh-coupling algorithm. Together, these methods allow for simultaneous segmentation of arbitrarily shaped adjacent objects. This chapter presents applications of coupled multi-object GraphCuts to selected structures of the musculoskeletal system in joint regions, namely bones and cartilage.

In this work, we build upon SSM-based segmentation, as introduced in Section 1.2.1. We describe the respective methodology in Section 3.1.

This chapter describes two pipelines that employ deformations of volumetric grids as described in Chapter 5.

In Chapter 4 we propose ODDS, a method that can handle free deformations of highly curved structures by allowing omnidirectional displacements for all vertices of a surface mesh during mesh deformation.

In this work we present a novel technique we term active graph matching, which integrates the popular active shape model into a sparse graph matching problem. This way we are able to combine the benefits of a global, statistical deformation model with the benefits of a local deformation model in form of a second-order random field. We present a new iterative energy minimization technique which achieves empirically good results. This enables us to exceed state-of-the art results for the task of annotating nuclei in 3D microscopic images of C. elegans. Furthermore with the help of the generalized Hough transform we are able to jointly segment and annotate a large set of nuclei in a fully automatic fashion for the first time.

Level-set methods are commonly used to segment regions of interest within images or volumes. These tasks usually involve a high number of operations. GPUs nowadays feature high computation and data throughput capabilities. In this work we present two GPU implementations of the level-set-based segmentation method called Fast Two Cycle. Our solutions partition the computational domain in tiles that can be processed in parallel. The original algorithm is adapted to the special features of the GPU, and performance is optimized by keeping a record of the tiles that require processing at any given time. We have tested our implementations with a set of 3D CT images of brain vessels and we show that we can obtain competitive results using commodity hardware.

Deformable surface models are often represented as triangular meshes in image segmentation applications. For a fast and easily regularized deformation onto the target object boundary, the vertices of the mesh are commonly moved along line segments (typically surface normals). However, in case of high mesh curvature, these lines may not intersect with the target boundary at all. Consequently, certain deformations cannot be achieved. We propose omnidirectional displacements for deformable surfaces (ODDS) to overcome this limitation. ODDS allow each vertex to move not only along a line segment but within the volumetric inside of a surrounding sphere, and achieve globally optimal deformations subject to local regularization constraints. However, allowing a ball-shaped instead of a linear range of motion per vertex significantly increases runtime and memory. To alleviate this drawback, we propose a hybrid approach, fastODDS, with improved runtime and reduced memory requirements. Furthermore, fastODDS can also cope with simultaneous segmentation of multiple objects. We show the theoretical benefits of ODDS with experiments on synthetic data, and evaluate ODDS and fastODDS quantitatively on clinical image data of the mandible and the hip bones. There, we assess both the global segmentation accuracy as well as local accuracy in high curvature regions, such as the tip-shaped mandibular coronoid processes and the ridge-shaped acetabular rims of the hip bones.

We propose a fully automatic method for tooth detection and classification in CT or cone-beam CT image data. First we compute an accurate segmentation of the maxilla bone. Based on this segmentation, our method computes a complete and optimal separation of the row of teeth into 16 subregions and classifies the resulting regions as existing or missing teeth. This serves as a prerequisite for further individual tooth segmentation. We show the robustness of our approach by providing extensive validation on 43 clinical head CT scans.

In this work we present an articulated statistical shape model (ASSM) of the human knee. The model incorporates statistical shape variation plus explicit degrees of freedom that model physiological joint motion. We also present a strategy for segmentation of the knee joint from medical image data. We show the potential of the model via an evaluation on a set of 40 clinical MRI datasets with manual expert segmentations available.

The exact localization of the mandibular nerve with respect to the bone is important for applications in dental implantology and maxillofacial surgery. Cone beam computed tomography (CBCT), often also called digital volume tomography (DVT), is increasingly utilized in maxillofacial or dental imaging. Compared to conventional CT, however, soft tissue discrimination is worse due to a reduced dose. Thus, small structures like the alveolar nerves are even harder recognizable within the image data. We show that it is nonetheless possible to accurately reconstruct the 3D bone surface and the course of the nerve in a fully automatic fashion, with a method that is based on a combined statistical shape model of the nerve and the bone and a Dijkstra-based optimization procedure. Our method has been validated on 106 clinical datasets: the average reconstruction error for the bone is 0.5 +/- 0.1 mm, and the nerve can be detected with an average error of 1.0 +/- 0.6 mm.

This work presents three different methods for automatic detection of anatomical landmarks in CT data, namely for the left and right anterior superior iliac spines and the pubic symphysis. The methods exhibit different degrees of generality in terms of portability to other anatomical landmarks and require a different amount of training data. The first method is problem-specific and is based on the convex hull of the pelvis. Method two is a more generic approach based on a statistical shape model including the landmarks of interest for every training shape. With our third method we present the most generic approach, where only a small set of training landmarks is required. Those landmarks are transferred to the patient specific geometry based on Mean Value Coordinates (MVCs). The methods work on surfaces of the pelvis that need to be extracted beforehand. We perform this geometry reconstruction with our previously introduced fully automatic segmentation framework for the pelvic bones. With a focus on the accuracy of our novel MVC-based approach, we evaluate and compare our methods on 100 clinical CT datasets, for which gold standard landmarks were defined manually by multiple observers.

For biomechanical simulations, the segmentation of multiple adjacent anatomical structures from medical image data is often required. If adjacent structures are hardly distinguishable in image data, automatic segmentation methods for single structures in general do not yield sufficiently accurate results. To improve segmentation accuracy in these cases, knowledge about adjacent structures must be exploited. Optimal graph searching based on deformable surface models allows for a simultaneous segmentation of multiple adjacent objects. However, this method requires a correspondence relation between vertices of adjacent surface meshes. Line segments, each containing two corresponding vertices, may then serve as shared displacement directions in the segmentation process. The problem is how to define suitable correspondences on arbitrary surfaces. In this paper we propose a scheme for constructing a correspondence relation in adjacent regions of two arbitrary surfaces. When applying the thus generated shared displacement directions in segmentation with deformable surfaces, overlap of the surfaces is guaranteed not to occur. We show correspondence relations for regions on a femoral head and acetabulum and other adjacent structures, as well as preliminary segmentation results obtained by a graph cut algorithm.

In this paper we propose a framework for fully automatic, robust and accurate segmentation of the human pelvis and proximal femur in CT data. We propose a composite statistical shape model of femur and pelvis with a flexible hip joint, for which we extend the common definition of statistical shape models as well as the common strategy for their adaptation. We do not analyze the joint flexibility statistically, but model it explicitly by rotational parameters describing the bent in a ball-and-socket joint. A leave-one-out evaluation on 50 CT volumes shows that image driven adaptation of our composite shape model robustly produces accurate segmentations of both proximal femur and pelvis. As a second contribution, we evaluate a fine grain multi-object segmentation method based on graph optimization. It relies on accurate initializations of femur and pelvis, which our composite shape model can generate. Simultaneous optimization of both femur and pelvis yields more accurate results than separate optimizations of each structure. Shape model adaptation and graph based optimization are embedded in a fully automatic framework.