In the era of modern computer vision (CV), deep learning (DL)-based methods have gained significant importance due to their remarkable performance in various applications such as object detection, recognition, segmentation, classification, and restoration in real-world scenarios. Currently, most of the existing survey papers for DL on CV applications provide adequate descriptions of a few methods used for only one or few task(s) in CV, such as object detection, recognition, and segmentation. The success of DL architectures relies profoundly on the effectiveness of loss functions used, appropriate for both the task and DL architecture being employed, especially for CV applications. This paper presents a comprehensive survey of various categories of loss functions used in several computer vision tasks. We have categorized them separately based on the application tasks being addressed, learning paradigms as well as the nature of their metric structure. Three categorized tables containing mathematical expressions of loss functions are cross-linked to extract better knowledge for any sub-group of loss functions. Organization of this survey will benefit researchers/developers to gain knowledge, identify, and use proper cost/energy functions for training DL models with architectural variants, for targeted application(s). Also, an organized list of loss functions, popularly used in the recent past, immensely benefits in the design of novel cost functions for DL. Our motivation and associated contributions stems from the above intentions and issues. Furthermore, we conduct ablation studies with a few widely used effective loss functions applied to address two prominent CV tasks, to illustrate the utility of this survey.

Brain reference atlases are essential for neuroscience experiments and data integration. However, histological atlases of the mouse brain, crucial in biomedical research, have not kept pace. Autofluorescence-based volumetric brain atlases are increasingly used but lack microscopic histological contrast, cytoarchitectonic information, corresponding MRI datasets, and often have truncated brainstems. Here, we present a multimodal, multiscale atlas of the laboratory mouse brain and head. The new reference brains include the whole head with consecutive Nissl and myelin serial section histology in three planes of section with 0.46 µm in-plane resolution, including intact brainstem, cranial nerves, and associated sensors and musculature. We provide reassembled histological volumes with 20mu isotropic resolution in stereotactic coordinates, determined using co-registered in vivo MRI and CT. In addition to conventional MRI contrasts, we provide diffusion MRI-based in vivo and ex vivo microstructural information, adding a valuable co-registered contrast modality that bridges MRI with cell-resolution histological data. We shift emphasis from compartmental annotations to stereotactic coordinates in the reference brains, offering a basis for evolving annotations over time and resolving conflicting neuroanatomical judgments by different experts. This new reference atlas facilitates integration of molecular cell type data and regional connectivity, serves as a model for similar atlases in other species, and sets a precedent for preserving extra-cranial nervous system structures. ### Competing Interest Statement The authors have declared no competing interest. NIHNIH, , MH114821

The human brain is believed to contain a full complement of neurons by the time of birth together with a substantial amount of the connectivity architecture, even though a significant amount of growth occurs postnatally. The developmental process leading to this outcome is not well understood in humans in comparison with model organisms. Previous magnetic resonance imaging (MRI) studies give three-dimensional coverage but not cellular resolution. In contrast, sparsely sampled histological or spatial omics analyses have provided cellular resolution but not dense whole brain coverage. To address the unmet need to provide a quantitative spatiotemporal map of developing human brain at cellular resolution, we leveraged tape-transfer assisted serial section histology to obtain contiguous histological series and unbiased imaging with dense coverage. Interleaved 20μ thick Nissl and H&E series and MRI volumes are co-registered into multimodal reference volumes with 60μ isotropic resolution, together with atlas annotations and a stereotactic coordinate system based on skull landmarks. The histological atlas volumes have significantly more contrast and texture than the MRI volumes. We computationally detect cells brain-wide to obtain quantitative characterization of the cytoarchitecture of the developing brain at 13-14 and 20-21 gestational weeks, providing the first comprehensive regional cell counts and characterizing the differential growth of the different brain compartments. Morphological characteristics permit segmentation of cell types from histology. We detected and quantified brain-wide distribution of mitotic figures representing dividing cells, providing an unprecedented spatiotemporal atlas of proliferative dynamics in the developing human brain. Further, we characterized the abundance and distribution of Cajal-Retzius cells, a transient cell population that plays essential roles in organizing glutamatergic cortical neurons into layers. Together, our study provides an unprecedented quantitative window into the developing human brain and the reference volumes and coordinate space should be useful for integrating spatial omics data sets with dense histological context. ### Competing Interest Statement The authors have declared no competing interest.

To understand biological intelligence we need to map neuronal networks in vertebrate brains. Mapping mesoscale neural circuitry is done using injections of tracers that label groups of neurons whose axons project to different brain regions. Since many neurons are labeled, it is difficult to follow individual axons. Previous approaches have instead quantified the regional projections using the total label intensity within a region. However, such a quantification is not biologically meaningful. We propose a new approach better connected to the underlying neurons by skeletonizing labeled axon fragments and then estimating a volumetric length density. Our approach uses a combination of deep nets and the Discrete Morse (DM) technique from computational topology. This technique takes into account nonlocal connectivity information and therefore provides noise-robustness. We demonstrate the utility and scalability of the approach on whole-brain tracer injected data. We also define and illustrate an information theoretic measure that quantifies the additional information obtained, compared to the skeletonized tracer injection fragments, when individual axon morphologies are available. Our approach is the first application of the DM technique to computational neuroanatomy. It can help bridge between single-axon skeletons and tracer injections, two important data types in mapping neural networks in vertebrates.

A current focus in neuroscience is to map neuronal cell types in whole vertebrate brains using different imaging modalities. Mapping modern molecular and anatomical datasets into a common atlas includes challenges that existing workflows do not adequately address: multimodal signals, missing data or non reference signals, and quantification of individual variation. Our solution implements a generative model describing the likelihood of data given a sequence of transforms of an atlas, and a maximum a posteriori estimation framework. Our approach allows composition of mappings across chains of datasets rather than only pairs, and computes metrics for geometric quantification. We study a range of datasets (in/ex-vivo MRI, STP and fMOST, 2D serial histology, snRNAseq prepared tissue), quantifying cell density and geometric fluctuations across covariates, and reveal that individual variation is often greater than differences due to tissue processing techniques. We provide open source code, dataset standards, and a web interface. This establishes a quantitative workflow for unifying multi-modal whole-brain images in an atlas framework, validated using mouse datasets, enabling large scale integration of datasets essential to modern neuroscience.

A current focus of research in neuroscience is to enumerate, map and annotate neuronal cell types in whole vertebrate brains using different modalities of data acquisition. Mapping these molecular and anatomical datasets into a common reference space remains a key challenge. While several brain-to-atlas mapping workflows exist, they do not adequately address challenges of modern high throughput neuroimaging, including multimodal and multiscale signals, missing data or non reference signals, and geometric quantification of individual variation. Our solution is to implement a generative statistical model that describes the likelihood of imaging data given a sequence of transforms of an atlas image, and a framework for maximum a posteriori estimation of unknown parameters capturing the issues listed above. The key idea in our approach is to minimize the difference between synthetic image volumes and real data over these parameter. Rather than merely using mappings as a "normalization" step, we implement tools for using their local metric changes as an opportunity for geometric quantification of technical and biological sources of variation in an unprecedented manner. While the framework is used to compute pairwise mappings, our approach particularly allows for easy compositions across chains of multimodality datasets. We apply these methods across a broad range of datasets including various combinations of in-vivo and ex-vivo MRI, 3D STP and fMOST data sets, 2D serial histology sections, and brains processed for snRNAseq with tissue partially removed. We show biological utility by quantifying cell density and diffeomorphic characterization of brain shape fluctuations across biological covariates. We note that the magnitude of individual variation is often greater than differences between different sample preparation techniques. To facilitate community accessibility, we implement our algorithm as open source, include a web based framework, and implement input and output dataset standards. Our work establishes a quantitative, scalable and streamlined workflow for unifying a broad spectrum of multi-modal whole-brain light microscopic data volumes into a coordinate-based atlas framework. This work enables large scale integration of whole brain data sets that are essential in modern neuroscience. ### Competing Interest Statement The authors have declared no competing interest.

A current focus of research in neuroscience is to enumerate, map and annotate neuronal cell types in whole vertebrate brains using different modalities of data acquisition. Mapping these molecular and anatomical datasets into a common reference space remains a key challenge. While several brain-to-atlas mapping workflows exist, they do not adequately address challenges of modern high throughput neuroimaging, including multimodal and multiscale signals, missing data or non reference signals, and geometric quantification of individual variation. Our solution is to implement a generative statistical model that describes the likelihood of imaging data given a sequence of transforms of an atlas image, and a framework for maximum a posteriori estimation of unknown parameters capturing the issues listed above. The key idea in our approach is to minimize the difference between synthetic image volumes and real data over these parameter. Rather than merely using mappings as a “normalization” step, we implement tools for using their local metric changes as an opportunity for geometric quantification of technical and biological sources of variation in an unprecedented manner. While the framework is used to compute pairwise mappings, our approach particularly allows for easy compositions across chains of multimodality datasets. We apply these methods across a broad range of datasets including various combinations of in-vivo and ex-vivo MRI, 3D STP and fMOST data sets, 2D serial histology sections, and brains processed for snRNAseq with tissue partially removed. We show biological utility by quantifying cell density and diffeomorphic characterization of brain shape fluctuations across biological covariates. We note that the magnitude of individual variation is often greater than differences between different sample preparation techniques. To facilitate community accessibility, we implement our algorithm as open source, include a web based framework, and implement input and output dataset standards. Our work establishes a quantitative, scalable and streamlined workflow for unifying a broad spectrum of multi-modal whole-brain light microscopic data volumes into a coordinate-based atlas framework. This work enables large scale integration of whole brain data sets that are essential in modern neuroscience. ### Competing Interest Statement The authors have declared no competing interest.

Characterizing cellular diversity at different levels of biological organization and across data modalities is a prerequisite to understanding the function of cell types in the brain. Classification of neurons is also essential to manipulate cell types in controlled ways and to understand their variation and vulnerability in brain disorders. The BRAIN Initiative Cell Census Network (BICCN) is an integrated network of data-generating centers, data archives, and data standards developers, with the goal of systematic multimodal brain cell type profiling and characterization. Emphasis of the BICCN is on the whole mouse brain with demonstration of prototype feasibility for human and nonhuman primate (NHP) brains. Here, we provide a guide to the cellular and spatial approaches employed by the BICCN, and to accessing and using these data and extensive resources, including the BRAIN Cell Data Center (BCDC), which serves to manage and integrate data across the ecosystem. We illustrate the power of the BICCN data ecosystem through vignettes highlighting several BICCN analysis and visualization tools. Finally, we present emerging standards that have been developed or adopted toward Findable, Accessible, Interoperable, and Reusable (FAIR) neuroscience. The combined BICCN ecosystem provides a comprehensive resource for the exploration and analysis of cell types in the brain.

An essential step toward understanding brain function is to establish a structural framework with cellular resolution on which multi-scale datasets spanning molecules, cells, circuits and systems can be integrated and interpreted1. Here, as part of the collaborative Brain Initiative Cell Census Network (BICCN), we derive a comprehensive cell type-based anatomical description of one exemplar brain structure, the mouse primary motor cortex, upper limb area (MOp-ul). Using genetic and viral labelling, barcoded anatomy resolved by sequencing, single-neuron reconstruction, whole-brain imaging and cloud-based neuroinformatics tools, we delineated the MOp-ul in 3D and refined its sublaminar organization. We defined around two dozen projection neuron types in the MOp-ul and derived an input-output wiring diagram, which will facilitate future analyses of motor control circuitry across molecular, cellular and system levels. This work provides a roadmap towards a comprehensive cellular-resolution description of mammalian brain architecture.

Rodrigo Muñoz-Castañeda* Brian Zingg*†, Katherine S. Matho*, Quanxin Wang*, Xiaoyin Chen*, Nicholas N. Foster†, Arun Narasimhan, Anan Li, Karla E. Hirokawa, Bingxing Huo, Samik Bannerjee, Laura Korobkova, Chris Sin Park, Young-Gyun Park, Michael S. Bienkowski, Uree Chon, Diek W. Wheeler, Xiangning Li, Yun Wang, Kathleen Kelly, Xu An, Sarojini M. Attili, Ian Bowman† , Anastasiia Bludova, Ali Cetin, Liya Ding, Rhonda Drewes, Florence D’Orazi, Corey Elowsky, Stephan Fischer, William Galbavy, Lei Gao†, Jesse Gillis, Peter A. Groblewski, Lin Gou†, Joel D. Hahn, Joshua T. Hatfield, Houri Hintiryan†, Jason Huang, Hideki Kondo, Xiuli Kuang, Philip Lesnar, Xu Li, Yaoyao Li, Mengkuan Lin, Lijuan Liu, Darrick Lo†, Judith Mizrachi, Stephanie Mok, Maitham Naeemi, Philip R. Nicovich, Ramesh Palaniswamy, Jason Palmer, Xiaoli Qi, Elise Shen, Yu-Chi Sun, Huizhong Tao, Wayne Wakemen, Yimin Wang, Peng Xie, Shenqin Yao, Jin Yuan , Muye Zhu†, Lydia Ng, Li I. Zhang, Byung Kook Lim, Michael Hawrylycz, Hui Gong, James C. Gee, Yongsoo Kim, Hanchuan Peng, Kwanghun Chuang, X William Yang, Qingming Luo, Partha P. Mitra, Anthony M. Zador, Hongkui Zeng, Giorgio A. Ascoli, Z Josh Huang, Pavel Osten, Julie A. Harris, Hong-Wei Dong†

Certain facial parts are salient (unique) in appearance, which substantially contribute to the holistic recognition of a subject. Occlusion of these salient parts deteriorates the performance of face recognition algorithms. In this paper, we propose a generative model to reconstruct the missing parts of the face which are under occlusion. The proposed generative model (SD-GAN) reconstructs a face preserving the illumination variation and identity of the face. A novel adversarial training algorithm has been designed for a bimodal mutually exclusive Generative Adversarial Network (GAN) model, for faster convergence. A novel adversarial "structural" loss function is also proposed, comprising of two components: a holistic and a local loss, characterized by SSIM and patch-wise MSE. Ablation studies on real and synthetically occluded face datasets reveal that our proposed technique outperforms the competing methods by a considerable margin, even for boosting the performance of Face Recognition.

Neuroscientific data analysis has traditionally relied on linear algebra and stochastic process theory. However, the tree-like shapes of neurons cannot be described easily as points in a vector space (the subtraction of two neuronal shapes is not a meaningful operation), and methods from computational topology are better suited to their analysis. Here we introduce methods from Discrete Morse (DM) Theory to extract the tree-skeletons of individual neurons from volumetric brain image data, and to summarize collections of neurons labelled by tracer injections. Since individual neurons are topologically trees, it is sensible to summarize the collection of neurons using a consensus tree-shape that provides a richer information summary than the traditional regional ‘connectivity matrix’ approach. The conceptually elegant DM approach lacks hand-tuned parameters and captures global properties of the data as opposed to previous approaches which are inherently local. For individual skeletonization of sparsely labelled neurons we obtain substantial performance gains over state-of-the-art non-topological methods (over 10% improvements in precision and faster proofreading). The consensus-tree summary of tracer injections incorporates the regional connectivity matrix information, but in addition captures the collective collateral branching patterns of the set of neurons connected to the injection site, and provides a bridge between single-neuron morphology and tracer-injection data.

Understanding of neuronal circuitry at cellular resolution within the brain has relied on tract tracing methods which involve careful observation and interpretation by experienced neuroscientists. With recent developments in imaging and digitization, this approach is no longer feasible with the large scale (terabyte to petabyte range) images. Machine learning based techniques, using deep networks, provide an efficient alternative to the problem. However, these methods rely on very large volumes of annotated images for training and have error rates that are too high for scientific data analysis, and thus requires a significant volume of human-in-the-loop proofreading. Here we introduce a hybrid architecture combining prior structure in the form of topological data analysis methods, based on discrete Morse theory, with the best-in-class deep-net architectures for the neuronal connectivity analysis. We show significant performance gains using our hybrid architecture on detection of topological structure (e.g. connectivity of neuronal processes and local intensity maxima on axons corresponding to synaptic swellings) with precision/recall close to 90% compared with human observers. We have adapted our architecture to a high performance pipeline capable of semantic segmentation of light microscopic whole-brain image data into a hierarchy of neuronal compartments. We expect that the hybrid architecture incorporating discrete Morse techniques into deep nets will generalize to other data domains.

Pose-Invariant Face Recognition (PIFR) has been a serious challenge in the general field of face recognition (FR). The performance of face recognition algorithms deteriorate due to various degradations such as pose, illuminaton, occlusions, blur, noise, aliasing, etc. In this paper, we deal with the problem of 3D pose variation of a face. for that we design and propose PosIX Generative Adversarial Network (PosIX-GAN) that has been trained to generate a set of nice (high quality) face images with 9 different pose variations, when provided with a face image in any arbitrary pose as input. The discriminator of the GAN has also been trained to perform the task of face recognition along with the job of discriminating between real and generated (fake) images. Results when evaluated using two benchmark datasets, reveal the superior performance of PosIX-GAN over state-of-the-art shallow as well as deep learning methods.

Introduction: Historically, patients with rotator cuff arthropathy had limited reconstructive options. The early generations of reverse total shoulder arthroplasty (rTSA) designs had increased failure rates due to loosening of glenoid baseplates secondary to excessive torques. In 1985, Paul Grammont introduced a prosthetic design changing the center of rotation that addressed this major complication. The Grammont principles remain the foundation of modern reverse total shoulder prostheses, although the original design has undergone several adaptations. We reviewed here the various aspects of prosthetic designs including baseplates, glenospheres, humeral components, and polyethylene bearing interfaces. Areas covered: We discuss the evolution, biomechanics, prosthetic options, and future direction for rTSA. A literature search using the PubMed database including review articles, biomechanical studies, and clinical trials pertaining to rTSA prothesis and outcomes. Expert commentary: Despite an expansion in the understanding of the biomechanics of the rotator cuff deficient shoulder and its effect on the reverse total shoulder prostheses, Grammont principles remain the foundation of contemporary rTSA designs. Further clinical studies are needed to assess how modern prosthetic modifications effect clinical and radiographic outcomes. Additionally, implants are being used in younger individuals with expanded indications, therefore, close clinical monitoring is needed to better evaluate their prosthetic longevity.

There is a need in modern neuroscience for accurate and automated image processing techniques for analyzing the large volume of neuroanatomical data. For e.g., the use of light microscopy to image whole mouse brains in a mesoscopic scale produces individual neuroanatomical data volumes in the TerraByte range. A fundamental task involves the detection and quantification of objects of a given type, e.g. neuronal nuclei or somata, in whole mouse brains. Traditionally this quantification is performed by human visual inspection with high accuracy, that is not scalable. When state-of-the-art CNN and SVM-based methods are used to solve this classification problem, they achieve accuracy levels between 85-92%. However, higher rates of precision and recall, close to that of humans are necessary. In this paper, we describe an unsupervised, iterative algorithm, which provides a high close to human performance for a specific problem of broad interest, i.e. detection of Green Fluorescent Protein labeled nuclei in whole mouse brains. The algorithm judiciously combines classical computer vision (CV) techniques and is focused on the complex problem of decomposing strong overlapped objects (nuclei). Our proposed iterative method uses features detected on ridge lines over distance transformation and an arc based iterative spatial-filling method to solve the problem. We demonstrate our results on two whole mouse brain data sets of Gigabyte resolution and compare it with manual annotation of the brains. Our results show that an aptly designed CV algorithm with classical feature extractors, when tailored to this problem of interest, achieves near-ideal humanlike performance. Quantitative analysis, when compared with the manually annotated ground truth, reveals that our approach performs better on whole mouse brain scans than general purpose machine learning (including deep CNN) methods.

Face Recognition (FR) using Convolutional Neural Network (CNN) based models have achieved considerable success in constrained environments. They however fail to perform well in unconstrained scenarios, especially when the images are captured using surveillance cameras. These probe samples suffer from degradations such as noise, poor illumination, low resolution, blur as well as aliasing, when compared to the rich training (gallery) set, comprising mostly of mugshot images captured in laboratory settings. These images in the training (gallery) set are crisp and have high contrast, compared to the probe samples. To cope with this scenario, we propose a novel dual-pathway generative adversarial network (DP-GAN) which maps low resolution images captured using surveillance camera into their corresponding high resolution images, which are gallery-like, using a novel combination of multi-scale reconstruction and Jensen-Shannon divergence based loss. These images thus obtained are then used to train a deep domain adaptation (deep-DA) network to perform the task of FR. The proposed network achieves superior results (>90%) on four benchmark surveillance face datasets, evident from the rank-1 recognition rates when compared with recent state-of-the-art CNN-based techniques.

Learning based on convolutional neural networks (CNNs) or deep learning has been a major research area with applications in face recognition (FR). However, performances of algorithms designed for FR are unsatisfactory when surveillance conditions severely degrade the test probes. The work presented in this paper has three contributions. First, it proposes a novel adaptive-CNN architecture of deep learning refurbished for domain adaptation (DA), to overcome the difference in feature distributions between the gallery and probe samples. The proposed architecture consists of three components: feature (FM), adaptive (AM) and classification (CM) modules. Secondly, a novel 2-stage algorithm for Mutually Exclusive Training (2-MET) based on stochastic gradient descent, has been proposed. The final stage of training in 2-MET freezes the layers of the FM and CM, while updating (tuning) only the parameters of the AM using a few probe (as target) samples. This helps the proposed deep-DA CNN to bridge the disparities in the distributions of the gallery and probe samples, resulting in enhanced domain-invariant representation for efficient deep-DA learning and classification. The third contribution comes from rigorous experimentations performed on three benchmark real-world surveillance face datasets with various kinds of degradations. This reveals the superior performance of the proposed adaptive-CNN architecture with 2-MET training, using Rank-1 recognition rates and ROC and CMC metrics, over many recent state-of-the-art techniques of CNN and DA.