We present a method combining affinity prediction with region agglomeration, which improves significantly upon the state of the art of neuron segmentation from electron microscopy (EM) in accuracy and scalability. Our method consists of a 3D U-Net, trained to predict affinities between voxels, followed by iterative region agglomeration. We train using a structured loss based on Malis, encouraging topologically correct segmentations obtained from affinity thresholding. Our extension consists of two parts: First, we present a quasi-linear method to compute the loss gradient, improving over the original quadratic algorithm. Second, we compute the gradient in two separate passes to avoid spurious gradient contributions in early training stages. Our predictions are accurate enough that simple learning-free percentile-based agglomeration outperforms more involved methods used earlier on inferior predictions. We present results on three diverse EM datasets, achieving relative improvements over previous results of 27, 15, and 250 percent. Our findings suggest that a single method can be applied to both nearly isotropic block-face EM data and anisotropic serial sectioned EM data. The runtime of our method scales linearly with the size of the volume and achieves a throughput of $\sim$∼ 2.6 seconds per megavoxel, qualifying our method for the processing of very large datasets.

MOTIVATION: A significant focus of biological research is to understand the development, organization and function of tissues. A particularly productive area of study is on single layer epithelial tissues in which the adherence junctions of cells form a 2D manifold that is fluorescently labeled. Given the size of the tissue, a microscope must collect a mosaic of overlapping 3D stacks encompassing the stained surface. Downstream interpretation is greatly simplified by preprocessing such a dataset as follows: (i) extracting and mapping the stained manifold in each stack into a single 2D projection plane, (ii) correcting uneven illumination artifacts, (iii) stitching the mosaic planes into a single, large 2D image and (iv) adjusting the contrast. RESULTS: We have developed PreMosa, an efficient, fully automatic pipeline to perform the four preprocessing tasks above resulting in a single 2D image of the stained manifold across which contrast is optimized and illumination is even. Notable features are as follows. First, the 2D projection step employs a specially developed algorithm that actually finds the manifold in the stack based on maximizing contrast, intensity and smoothness. Second, the projection step comes first, implying all subsequent tasks are more rapidly solved in 2D. And last, the mosaic melding employs an algorithm that globally adjusts contrasts amongst the 2D tiles so as to produce a seamless, high-contrast image. We conclude with an evaluation using ground-truth datasets and present results on datasets from Drosophila melanogaster wings and Schmidtae mediterranea ciliary components. AVAILABILITY AND IMPLEMENTATION: PreMosa is available under https://cblasse.github.io/premosa. CONTACT: [email protected] or [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Motivation: Serial section microscopy is an established method for detailed anatomy reconstruction of biological specimen. During the last decade, high resolution electron microscopy (EM) of serial sections has become the de-facto standard for reconstruction of neural connectivity at ever increasing scales (EM connectomics). In serial section microscopy, the axial dimension of the volume is sampled by physically removing thin sections from the embedded specimen and subsequently imaging either the block-face or the section series. This process has limited precision leading to inhomogeneous non-planar sampling of the axial dimension of the volume which, in turn, results in distorted image volumes. This includes that section series may be collected and imaged in unknown order. Results: We developed methods to identify and correct these distortions through image-based signal analysis without any additional physical apparatus or measurements. We demonstrate the efficacy of our methods in proof of principle experiments and application to real world problems. Availability and Implementation: We made our work available as libraries for the ImageJ distribution Fiji and for deployment in a high performance parallel computing environment. Our sources are open and available at http://github.com/saalfeldlab/section-sort, http://github.com/saalfeldlab/z-spacing and http://github.com/saalfeldlab/z-spacing-spark. Contact: [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

Vol. 28 no. 22 2012, pages 3009–3011 doi:10.1093/bioinformatics/bts543 We regret that, due to a production error, the below references were incorrect in the original paper and should appear as below: Yoo,T.S. et al. (2002) Engineering and algorithm design for an image processing API: A technical report on ITK - the insight toolkit. In: Westwood, J. et al. (eds.) MMVR, pp. 586–592. Preibisch,S. et al. (2010) Into ImgLib—generic image processing in Java. In: Jahnen,A. and Moll,C. (eds.) ImageJ User and Developer Conference. Mondorf-les-Bains, Luxembourg. Rueden,C. et al. (2010) ImageJDev: Next generation ImageJ. In: Jahnen,A. and Moll,C. (eds.) ImageJ User and Developer Conference. Mondorf-les-Bains, Luxembourg. Matas,J. et al. (2002) Robust wide baseline stereo from maximally stable extremal regions. In: Marshall,D. and Rosin,P.L. (eds.) BMVC, Vol. 1, pp. 384–393. Nistér,D. and Stewénius,H. (2008) Linear time maximally stable extremal regions. In: Forsyth,D. et al. (eds.) ECCV, pp. 183–196.

SUMMARY: ImgLib2 is an open-source Java library for n-dimensional data representation and manipulation with focus on image processing. It aims at minimizing code duplication by cleanly separating pixel-algebra, data access and data representation in memory. Algorithms can be implemented for classes of pixel types and generic access patterns by which they become independent of the specific dimensionality, pixel type and data representation. ImgLib2 illustrates that an elegant high-level programming interface can be achieved without sacrificing performance. It provides efficient implementations of common data types, storage layouts and algorithms. It is the data model underlying ImageJ2, the KNIME Image Processing toolbox and an increasing number of Fiji-Plugins. AVAILABILITY: ImgLib2 is licensed under BSD. Documentation and source code are available at http://imglib2.net and in a public repository at https://github.com/imagej/imglib. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics Online. CONTACT: [email protected]

MOTIVATION: Tiled serial section Transmission Electron Microscopy (ssTEM) is increasingly used to describe high-resolution anatomy of large biological specimens. In particular in neurobiology, TEM is indispensable for analysis of synaptic connectivity in the brain. Registration of ssTEM image mosaics has to recover the 3D continuity and geometrical properties of the specimen in presence of various distortions that are applied to the tissue during sectioning, staining and imaging. These include staining artifacts, mechanical deformation, missing sections and the fact that structures may appear dissimilar in consecutive sections. RESULTS: We developed a fully automatic, non-rigid but as-rigid-as-possible registration method for large tiled serial section microscopy stacks. We use the Scale Invariant Feature Transform (SIFT) to identify corresponding landmarks within and across sections and globally optimize the pose of all tiles in terms of least square displacement of these landmark correspondences. We evaluate the precision of the approach using an artificially generated dataset designed to mimic the properties of TEM data. We demonstrate the performance of our method by registering an ssTEM dataset of the first instar larval brain of Drosophila melanogaster consisting of 6885 images. AVAILABILITY: This method is implemented as part of the open source software TrakEM2 (http://www.ini.uzh.ch/~acardona/trakem2.html) and distributed through the Fiji project (http://pacific.mpi-cbg.de).

MOTIVATION: Modern anatomical and developmental studies often require high-resolution imaging of large specimens in three dimensions (3D). Confocal microscopy produces high-resolution 3D images, but is limited by a relatively small field of view compared with the size of large biological specimens. Therefore, motorized stages that move the sample are used to create a tiled scan of the whole specimen. The physical coordinates provided by the microscope stage are not precise enough to allow direct reconstruction (Stitching) of the whole image from individual image stacks. RESULTS: To optimally stitch a large collection of 3D confocal images, we developed a method that, based on the Fourier Shift Theorem, computes all possible translations between pairs of 3D images, yielding the best overlap in terms of the cross-correlation measure and subsequently finds the globally optimal configuration of the whole group of 3D images. This method avoids the propagation of errors by consecutive registration steps. Additionally, to compensate the brightness differences between tiles, we apply a smooth, non-linear intensity transition between the overlapping images. Our stitching approach is fast, works on 2D and 3D images, and for small image sets does not require prior knowledge about the tile configuration. AVAILABILITY: The implementation of this method is available as an ImageJ plugin distributed as a part of the Fiji project (Fiji is just ImageJ: http://pacific.mpi-cbg.de/).

SUMMARY: High-resolution, three-dimensional (3D) imaging of large biological specimens generates massive image datasets that are difficult to navigate, annotate and share effectively. Inspired by online mapping applications like GoogleMaps, we developed a decentralized web interface that allows seamless navigation of arbitrarily large image stacks. Our interface provides means for online, collaborative annotation of the biological image data and seamless sharing of regions of interest by bookmarking. The CATMAID interface enables synchronized navigation through multiple registered datasets even at vastly different scales such as in comparisons between optical and electron microscopy. AVAILABILITY: http://fly.mpi-cbg.de/catmaid.