While recent machine learning research has revealed connections between deep generative models such as VAEs and rate-distortion losses used in learned compression, most of this work has focused on images. In a similar spirit, we view recently proposed neural video coding algorithms through the lens of deep autoregressive and latent variable modeling. We present these codecs as instances of a generalized stochastic temporal autoregressive transform, and propose new avenues for further improvements inspired by normalizing flows and structured priors. We propose several architectures that yield state-of-the-art video compression performance on high-resolution video and discuss their tradeoffs and ablations. In particular, we propose (i) improved temporal autoregressive transforms, (ii) improved entropy models with structured and temporal dependencies, and (iii) variable bitrate versions of our algorithms. Since our improvements are compatible with a large class of existing models, we provide further evidence that the generative modeling viewpoint can advance the neural video coding field.

The growing complexity of spatial and structural information in 3D data makes data inspection and visualization a challenging task. We describe a method to create a planar embedding of 3D treelike structures using their skeleton representations. Our method maintains the original geometry, without overlaps, to the best extent possible, allowing exploration of the topology within a single view. We present a novel camera view generation method which maximizes the visible geometric attributes (segment shape and relative placement between segments). Camera views are created for individual segments and are used to determine local bending angles at each node by projecting them to 2D. The final embedding is generated by minimizing an energy function (the weights of which are user adjustable) based on branch length and the 2D angles, while avoiding intersections. The user can also interactively modify segment placement within the 2D embedding, and the overall embedding will update accordingly. A global to local interactive exploration is provided using hierarchical camera views that are created for subtrees within the structure. We evaluate our method both qualitatively and quantitatively and demonstrate our results by constructing planar visualizations of line data (traced neurons) and volume data (CT vascular and bronchial data).

In the clinic, resected tissue samples are stained with Hematoxylin-and-Eosin (H&E) and/or Immunhistochemistry (IHC) stains and presented to the pathologists on glass slides or as digital scans for diagnosis and assessment of disease progression. Cell-level quantification, e.g. in IHC protein expression scoring, can be extremely inefficient and subjective. We present DeepLIIF (https://deepliif.org), a first free online platform for efficient and reproducible IHC scoring. DeepLIIF outperforms current state-of-the-art approaches (relying on manual error-prone annotations) by virtually restaining clinical IHC slides with more informative multiplex immunofluorescence staining. Our DeepLIIF cloud-native platform supports (1) more than 150 proprietary/non-proprietary input formats via the Bio-Formats standard, (2) interactive adjustment, visualization, and downloading of the IHC quantification results and the accompanying restained images, (3) consumption of an exposed workflow API programmatically or through interactive plugins for open source whole slide image viewers such as QuPath/ImageJ, and (4) auto scaling to efficiently scale GPU resources based on user demand.

We present a review of predictive coding, from theoretical neuroscience, and variational autoencoders, from machine learning, identifying the common origin and mathematical framework underlying both areas. As each area is prominent within its respective field, more firmly connecting these areas could prove useful in the dialogue between neuroscience and machine learning. After reviewing each area, we discuss two possible correspondences implied by this perspective: cortical pyramidal dendrites as analogous to (nonlinear) deep networks and lateral inhibition as analogous to normalizing flows. These connections may provide new directions for further investigations in each field.

We present 3D virtual pancreatography (VP), a novel visualization procedure and application for non-invasive diagnosis and classification of pancreatic lesions, the precursors of pancreatic cancer. Currently, non-invasive screening of patients is performed through visual inspection of 2D axis-aligned CT images, though the relevant features are often not clearly visible nor automatically detected. VP is an end-to-end visual diagnosis system that includes: A machine learning based automatic segmentation of the pancreatic gland and the lesions, a semi-automatic approach to extract the primary pancreatic duct, a machine learning based automatic classification of lesions into four prominent types, and specialized 3D and 2D exploratory visualizations of the pancreas, lesions and surrounding anatomy. We combine volume rendering with pancreas- and lesion-centric visualizations and measurements for effective diagnosis. We designed VP through close collaboration and feedback from expert radiologists, and evaluated it on multiple real-world CT datasets with various pancreatic lesions and case studies examined by the expert radiologists.

Policy networks are a central feature of deep reinforcement learning (RL) algorithms for continuous control, enabling the estimation and sampling of high-value actions. From the variational inference perspective on RL, policy networks, when used with entropy or KL regularization, are a form of amortized optimization, optimizing network parameters rather than the policy distributions directly. However, direct amortized mappings can yield suboptimal policy estimates and restricted distributions, limiting performance and exploration. Given this perspective, we consider the more flexible class of iterative amortized optimizers. We demonstrate that the resulting technique, iterative amortized policy optimization, yields performance improvements over direct amortization on benchmark continuous control tasks.

Machine learning is a powerful and effective tool for medical image analysis to perform computer-aided diagnosis (CAD). Having great potential in improving the accuracy of a diagnosis, CAD systems are often analyzed in terms of the final accuracy, leading to a limited understanding of the internal decision process, impossibility to gain insights, and ultimately to skepticism from clinicians. We present a visual analytics approach to uncover the decision-making process of a CAD system for classifying pancreatic cystic lesions. This CAD algorithm consists of two distinct components: random forest (RF), which classifies a set of predefined features, including demographic features, and a convolutional neural network (CNN), which analyzes radiological (imaging) features of the lesions. We study the class probabilities generated by the RF and the semantical meaning of the features learned by the CNN. We also use an eye tracker to better understand which radiological features are particularly useful for a radiologist to make a diagnosis and to quantitatively compare with the features that lead the CNN to its final classification decision. Additionally, we evaluate the effects and benefits of supplying the CAD system with a case-based visual aid in a second-reader setting.

We present a visual analytics framework, CMed, for exploring medical image data annotations acquired from crowdsourcing. CMed can be used to visualize, classify, and filter crowdsourced clinical data based on a number of different metrics such as detection rate, logged events, and clustering of the annotations. CMed provides several interactive linked visualization components to analyze the crowd annotation results for a particular video and the associated workers. Additionally, all results of an individual worker can be inspected using multiple linked views in our CMed framework. We allow a crowdsourcing application analyst to observe patterns and gather insights into the crowdsourced medical data, helping him/her design future crowdsourcing applications for optimal output from the workers. We demonstrate the efficacy of our framework with two medical crowdsourcing studies: polyp detection in virtual colonoscopy videos and lung nodule detection in CT thin-slab maximum intensity projection videos. We also provide experts' feedback to show the effectiveness of our framework. Lastly, we share the lessons we learned from our framework with suggestions for integrating our framework into a clinical workflow.

Navigation is a major challenge in exploring data within immersive environments, especially of large omnidirectional spherical images. We propose a method of auto-scaling to allow users to navigate using teleportation within the safe boundary of their physical environment with different levels of focus. Our method combines physical navigation with virtual teleportation. We also propose a “peek then warp” behavior when using a zoom lens and evaluate our system in conjunction with different teleportation transitions, including a proposed transition for exploration of omnidirectional and 360-degree panoramic imagery, termed Envelop, wherein the destination view expands out from the zoom lens to completely envelop the user. In this work, we focus on visualizing and navigating large omnidirectional or panoramic images with application to GIS visualization as an inside-out omnidirectional image of the earth. We conducted two user studies to evaluate our techniques over a search and comparison task. Our results illustrate the advantages of our techniques for navigation and exploration of omnidirectional images in an immersive environment.

Virtual colonoscopy (VC) is a non-invasive screening tool for colorectal polyps which employs volume visualization of a colon model reconstructed from a CT scan of the patient's abdomen. We present an immersive analytics system for VC which enhances and improves the traditional desktop VC through the use of VR technologies. Our system, using a head-mounted display (HMD), includes all of the standard VC features, such as the volume rendered endoluminal fly-through, measurement tool, bookmark modes, electronic biopsy, and slice views. The use of VR immersion, stereo, and wider field of view and field of regard has a positive effect on polyp search and analysis tasks in our immersive VC system, a volumetric-based immersive analytics application. Navigation includes enhanced automatic speed and direction controls, based on the user's head orientation, in conjunction with physical navigation for exploration of local proximity. In order to accommodate the resolution and frame rate requirements for HMDs, new rendering techniques have been developed, including mesh-assisted volume raycasting and a novel lighting paradigm. Feedback and further suggestions from expert radiologists show the promise of our system for immersive analysis for VC and encourage new avenues for exploring the use of VR in visualization systems for medical diagnosis.

Inference models are a key component in scaling variational inference to deep latent variable models, most notably as encoder networks in variational auto-encoders (VAEs). By replacing conventional optimization-based inference with a learned model, inference is amortized over data examples and therefore more computationally efficient. However, standard inference models are restricted to direct mappings from data to approximate posterior estimates. The failure of these models to reach fully optimized approximate posterior estimates results in an amortization gap. We aim toward closing this gap by proposing iterative inference models, which learn to perform inference optimization through repeatedly encoding gradients. Our approach generalizes standard inference models in VAEs and provides insight into several empirical findings, including top-down inference techniques. We demonstrate the inference optimization capabilities of iterative inference models and show that they outperform standard inference models on several benchmark data sets of images and text.

We introduce the variational filtering EM algorithm, a simple, general-purpose method for performing variational inference in dynamical latent variable models using information from only past and present variables, i.e. filtering. The algorithm is derived from the variational objective in the filtering setting and consists of an optimization procedure at each time step. By performing each inference optimization procedure with an iterative amortized inference model, we obtain a computationally efficient implementation of the algorithm, which we call amortized variational filtering. We present experiments demonstrating that this general-purpose method improves inference performance across several recent deep dynamical latent variable models.

For many virtual reality applications, a pre-calculated fly-through path is the de facto standard navigation method. Such a path is convenient for users and ensures coverage of critical areas throughout the scene. Traditional applications use constant camera speed, allow for fully user-controlled manual speed adjustment, or use automatic speed adjustment based on heuristics from the scene. We introduce two novel methods for constrained path navigation and exploration in virtual environments which rely on the natural orientation of the user's head during scene exploration. Utilizing head tracking to obtain the user's area of focus, we perform automatic camera speed adjustment to allow for natural off-axis scene examination. We expand this to include automatic camera navigation along the pre-computed path, abrogating the need for any navigational inputs from the user. Our techniques are applicable for any scene with a pre-computed navigation path, including medical applications such as virtual colonoscopy, coronary fly-through, or virtual angioscopy, and graph navigation. We compare the traditional methods (constant speed and manual speed adjustment) and our two methods (automatic speed adjustment and automatic speed/direction control) to determine the effect of speed adjustment on system usability, mental load, performance, and user accuracy. Through this evaluation we observe the effect of automatic speed adjustment compared to traditional methods. We observed no negative impact from automatic navigation, and the users performed as well as with the manual navigation.

We present a method for registration and visualization of corresponding supine and prone virtual colonoscopy scans based on eigenfunction analysis and fold modeling. In virtual colonoscopy, CT scans are acquired with the patient in two positions, and their registration is desirable so that physicians can corroborate findings between scans. Our algorithm performs this registration efficiently through the use of Fiedler vector representation (the second eigenfunction of the Laplace-Beltrami operator). This representation is employed to first perform global registration of the two colon positions. The registration is then locally refined using the haustral folds, which are automatically segmented using the 3D level sets of the Fiedler vector. The use of Fiedler vectors and the segmented folds presents a precise way of visualizing corresponding regions across datasets and visual modalities. We present multiple methods of visualizing the results, including 2D flattened rendering and the corresponding 3D endoluminal views. The precise fold modeling is used to automatically find a suitable cut for the 2D flattening, which provides a less distorted visualization. Our approach is robust, and we demonstrate its efficiency and efficacy by showing matched views on both the 2D flattened colons and in the 3D endoluminal view. We analytically evaluate the results by measuring the distance between features on the registered colons, and we also assess our fold segmentation against 20 manually labeled datasets. We have compared our results analytically to previous methods, and have found our method to achieve superior results. We also prove the hot spots conjecture for modeling cylindrical topology using Fiedler vector representation, which allows our approach to be used for general cylindrical geometry modeling and feature extraction.

We present a novel method to create planar visualizations of treelike structures (e.g., blood vessels and airway trees) where the shape of the object is well preserved, allowing for easy recognition by users familiar with the structures. Based on the extracted skeleton within the treelike object, a radial planar embedding is first obtained such that there are no self-intersections of the skeleton which would have resulted in occlusions in the final view. An optimization procedure which adjusts the angular positions of the skeleton nodes is then used to reconstruct the shape as closely as possible to the original, according to a specified view plane, which thus preserves the global geometric context of the object. Using this shape recovered embedded skeleton, the object surface is then flattened to the plane without occlusions using harmonic mapping. The boundary of the mesh is adjusted during the flattening step to account for regions where the mesh is stretched over concavities. This parameterized surface can then be used either as a map for guidance during endoluminal navigation or directly for interrogation and decision making. Depth cues are provided with a grayscale border to aid in shape understanding. Examples are presented using bronchial trees, cranial and lower limb blood vessels, and upper aorta datasets, and the results are evaluated quantitatively and with a user study.

Abstract Prostate cancer is one of the most prevalent cancers among males, and the use of magnetic resonance imaging (MRI) has been suggested for its detection. A framework is presented for scoring and visualizing various MR data in an efficient and intuitive manner. A classification method is introduced where a cumulative score volume is created which takes into account each of three acquisition types. This score volume is integrated into a volume rendering framework which allows the user to view the prostate gland, the multi‐modal score values, and the surrounding anatomy. A visibility persistence mode is introduced to automatically avoid full occlusion of a selected score and indicate overlaps. The use of GPU‐accelerated multi‐modal single‐pass ray casting provides an interactive experience. User driven importance rendering allows the user to gain insight into the data and can assist in localization of the disease and treatment planning. We evaluate our results against pathology and radiologists'determinations.

When visualizing tubular 3D structures, external representations are often used for guidance and display, and such views in 2D can often contain occlusions. Virtual dissection methods have been proposed where the entire 3D structure can be mapped to the 2D plane, though these will lose context by straightening curved sections. We present a new method of creating maps of 3D tubular structures that yield a succinct view while preserving the overall geometric structure. Given a dominant view plane for the structure, its curve skeleton is first projected to a 2D skeleton. This 2D skeleton is adjusted to account for distortions in length, modified to remove intersections, and optimized to preserve the shape of the original 3D skeleton. Based on this shaped 2D skeleton, a boundary for the map of the object is obtained based on a slicing path through the structure and the radius around the skeleton. The sliced structure is conformally mapped to a rectangle and then deformed via harmonic mapping to match the boundary placement. This flattened map preserves the general geometric context of a 3D object in a 2D display, and rendering of this flattened map can be accomplished using volumetric ray casting. We have evaluated our method on real datasets of human colon models.

In virtual colonoscopy, CT scans are typically acquired with the patient in both supine (facing up) and prone (facing down) positions. The registration of these two scans is desirable so that the user can clarify situations or confirm polyp findings at a location in one scan with the same location in the other, thereby improving polyp detection rates and reducing false positives. However, this supine-prone registration is challenging because of the substantial distortions in the colon shape due to the patient's change in position. We present an efficient algorithm and framework for performing this registration through the use of conformal geometry to guarantee that the registration is a diffeomorphism (a one-to-one and onto mapping). The taeniae coli and colon flexures are automatically extracted for each supine and prone surface, employing the colon geometry. The two colon surfaces are then divided into several segments using the flexures, and each segment is cut along a taenia coli and conformally flattened to the rectangular domain using holomorphic differentials. The mean curvature is color encoded as texture images, from which feature points are automatically detected using graph cut segmentation, mathematic morphological operations, and principal component analysis. Corresponding feature points are found between supine and prone and are used to adjust the conformal flattening to be quasi-conformal, such that the features become aligned. We present multiple methods of visualizing our results, including 2D flattened rendering, corresponding 3D endoluminal views, and rendering of distortion measurements. We demonstrate the efficiency and efficacy of our registration method by illustrating matched views on both the 2D flattened colon images and in the 3D volume rendered colon endoluminal view. We analytically evaluate the correctness of the results by measuring the distance between features on the registered colons.