Hannes Nickisch

Zohaib Salahuddin, Matthias Lenga, Hannes Nickisch

We propose a deep learning-based automatic coronary artery tree centerline tracker (AuCoTrack) extending the vessel tracker by Wolterink (arXiv:1810.03143). A dual pathway Convolutional Neural Network (CNN) operating on multi-scale 3D inputs predicts the direction of the coronary arteries as well as the presence of a bifurcation. A similar multi-scale dual pathway 3D CNN is trained to identify coronary artery endpoints for terminating the tracking process. Two or more continuation directions are derived based on the bifurcation detection. The iterative tracker detects the entire left and right coronary artery trees based on only two ostium landmarks derived from a model-based segmentation of the heart. The 3D CNNs were trained on a proprietary dataset consisting of 43 CCTA scans. An average sensitivity of 87.1% and clinically relevant overlap of 89.1% was obtained relative to a refined manual segmentation. In addition, the MICCAI 2008 Coronary Artery Tracking Challenge (CAT08) training and test datasets were used to benchmark the algorithm and to assess its generalization. An average overlap of 93.6% and a clinically relevant overlap of 96.4% were obtained. The proposed method achieved better overlap scores than the current state-of-the-art automatic centerline extraction techniques on the CAT08 dataset with a vessel detection rate of 95%.

Ivo M. Baltruschat, Leonhard Steinmeister, Hannes Nickisch et al.

The aim is to evaluate whether smart worklist prioritization by artificial intelligence (AI) can optimize the radiology workflow and reduce report turnaround times (RTAT) for critical findings in chest radiographs (CXRs). Furthermore, we investigate a method to counteract the effect of false negative predictions by AI -- resulting in an extremely and dangerously long RTAT, as CXRs are sorted to the end of the worklist. We developed a simulation framework that models the current workflow at a university hospital by incorporating hospital specific CXR generation rates, reporting rates and pathology distribution. Using this, we simulated the standard worklist processing "first-in, first-out" (FIFO) and compared it with a worklist prioritization based on urgency. Examination prioritization was performed by the AI, classifying eight different pathological findings ranked in descending order of urgency: pneumothorax, pleural effusion, infiltrate, congestion, atelectasis, cardiomegaly, mass and foreign object. Furthermore, we introduced an upper limit for the maximum waiting time, after which the highest urgency is assigned to the examination. The average RTAT for all critical findings was significantly reduced in all Prioritization-simulations compared to the FIFO-simulation (e.g. pneumothorax: 35.6 min vs. 80.1 min; p $<0.0001$), while the maximum RTAT for most findings increased at the same time (e.g. pneumothorax: 1293 min vs 890 min; p $<0.0001$). Our "upper limit" substantially reduced the maximum RTAT all classes (e.g. pneumothorax: 979 min vs. 1293 min / 1178 min; p $<0.0001$). Our simulations demonstrate that smart worklist prioritization by AI can reduce the average RTAT for critical findings in CXRs while maintaining a small maximum RTAT as FIFO.

Moti Freiman, Hannes Nickisch, Holger Schmitt et al.

We introduce a functional for the learning of an optimal database for patch-based image segmentation with application to coronary lumen segmentation from coronary computed tomography angiography (CCTA) data. The proposed functional consists of fidelity, sparseness and robustness to small-variations terms and their associated weights. Existing work address database optimization by prototype selection aiming to optimize the database by either adding or removing prototypes according to a set of predefined rules. In contrast, we formulate the database optimization task as an energy minimization problem that can be solved using standard numerical tools. We apply the proposed database optimization functional to the task of optimizing a database for patch-base coronary lumen segmentation. Our experiments using the publicly available MICCAI 2012 coronary lumen segmentation challenge data show that optimizing the database using the proposed approach reduced database size by 96% while maintaining the same level of lumen segmentation accuracy. Moreover, we show that the optimized database yields an improved specificity of CCTA based fractional flow reserve (0.73 vs 0.7 for all lesions and 0.68 vs 0.65 for obstructive lesions) using a training set of 132 (76 obstructive) coronary lesions with invasively measured FFR as the reference.

Moti Freiman, Hannes Nickisch, Sven Prevrhal et al.

Purpose: The goal of this study was to assess the potential added benefit of accounting for partial volume effects (PVE) in an automatic coronary lumen segmentation algorithm from coronary computed tomography angiography (CCTA). Materials and methods: We assessed the potential added value of PVE integration as a part of the automatic coronary lumen segmentation algorithm by means of segmentation accuracy using the MICCAI 2012 challenge framework and by means of flow simulation overall accuracy, sensitivity, specificity, negative and positive predictive values and the receiver operated characteristic (ROC) area under the curve. We also evaluated the potential benefit of accounting for PVE in automatic segmentation for flow-simulation for lesions that were diagnosed as obstructive based on CCTA, which could have indicated a need for an invasive exam and revascularization. Results: Our segmentation algorithm improves the maximal surface distance error by ~39% compared to previously published method on the 18 datasets 50 from the MICCAI 2012 challenge with comparable Dice and mean surface distance. Results with and without accounting for PVE were comparable. In contrast, integrating PVE analysis into an automatic coronary lumen segmentation algorithm improved the flow simulation specificity from 0.6 to 0.68 with the same sensitivity of 0.83. Also, accounting for PVE improved the area under the ROC curve for detecting hemodynamically significant CAD from 0.76 to 0.8 compared to automatic segmentation without PVE analysis with invasive FFR threshold of 0.8 as the reference standard. The improvement in the AUC was statistically significant (N=76, Delong's test, p=0.012). Conclusion: Accounting for the partial volume effects in automatic coronary lumen segmentation algorithms has the potential to improve the accuracy of CCTA-based hemodynamic assessment of coronary artery lesions.

William Herlands, Daniel B. Neill, Hannes Nickisch et al.

Identifying changes in model parameters is fundamental in machine learning and statistics. However, standard changepoint models are limited in expressiveness, often addressing unidimensional problems and assuming instantaneous changes. We introduce change surfaces as a multidimensional and highly expressive generalization of changepoints. We provide a model-agnostic formalization of change surfaces, illustrating how they can provide variable, heterogeneous, and non-monotonic rates of change across multiple dimensions. Additionally, we show how change surfaces can be used for counterfactual prediction. As a concrete instantiation of the change surface framework, we develop Gaussian Process Change Surfaces (GPCS). We demonstrate counterfactual prediction with Bayesian posterior mean and credible sets, as well as massive scalability by introducing novel methods for additive non-separable kernels. Using two large spatio-temporal datasets we employ GPCS to discover and characterize complex changes that can provide scientific and policy relevant insights. Specifically, we analyze twentieth century measles incidence across the United States and discover previously unknown heterogeneous changes after the introduction of the measles vaccine. Additionally, we apply the model to requests for lead testing kits in New York City, discovering distinct spatial and demographic patterns.

Ivo M. Baltruschat, Leonhard Steinmeister, Harald Ittrich et al.

Chest radiography is the most common clinical examination type. To improve the quality of patient care and to reduce workload, methods for automatic pathology classification have been developed. In this contribution we investigate the usefulness of two advanced image pre-processing techniques, initially developed for image reading by radiologists, for the performance of Deep Learning methods. First, we use bone suppression, an algorithm to artificially remove the rib cage. Secondly, we employ an automatic lung field detection to crop the image to the lung area. Furthermore, we consider the combination of both in the context of an ensemble approach. In a five-times re-sampling scheme, we use Receiver Operating Characteristic (ROC) statistics to evaluate the effect of the pre-processing approaches. Using a Convolutional Neural Network (CNN), optimized for X-ray analysis, we achieve a good performance with respect to all pathologies on average. Superior results are obtained for selected pathologies when using pre-processing, i.e. for mass the area under the ROC curve increased by 9.95%. The ensemble with pre-processed trained models yields the best overall results.

Ivo M. Baltruschat, Hannes Nickisch, Michael Grass et al.

The increased availability of X-ray image archives (e.g. the ChestX-ray14 dataset from the NIH Clinical Center) has triggered a growing interest in deep learning techniques. To provide better insight into the different approaches, and their applications to chest X-ray classification, we investigate a powerful network architecture in detail: the ResNet-50. Building on prior work in this domain, we consider transfer learning with and without fine-tuning as well as the training of a dedicated X-ray network from scratch. To leverage the high spatial resolution of X-ray data, we also include an extended ResNet-50 architecture, and a network integrating non-image data (patient age, gender and acquisition type) in the classification process. In a concluding experiment, we also investigate multiple ResNet depths (i.e. ResNet-38 and ResNet-101). In a systematic evaluation, using 5-fold re-sampling and a multi-label loss function, we compare the performance of the different approaches for pathology classification by ROC statistics and analyze differences between the classifiers using rank correlation. Overall, we observe a considerable spread in the achieved performance and conclude that the X-ray-specific ResNet-38, integrating non-image data yields the best overall results. Furthermore, class activation maps are used to understand the classification process, and a detailed analysis of the impact of non-image features is provided.

Hannes Nickisch, Arno Solin, Alexander Grigorievskiy

We provide a comprehensive overview and tooling for GP modeling with non-Gaussian likelihoods using state space methods. The state space formulation allows for solving one-dimensional GP models in $\mathcal{O}(n)$ time and memory complexity. While existing literature has focused on the connection between GP regression and state space methods, the computational primitives allowing for inference using general likelihoods in combination with the Laplace approximation (LA), variational Bayes (VB), and assumed density filtering (ADF, a.k.a. single-sweep expectation propagation, EP) schemes has been largely overlooked. We present means of combining the efficient $\mathcal{O}(n)$ state space methodology with existing inference methods. We extend existing methods, and provide unifying code implementing all approaches.

Kun Dong, David Eriksson, Hannes Nickisch et al.

For applications as varied as Bayesian neural networks, determinantal point processes, elliptical graphical models, and kernel learning for Gaussian processes (GPs), one must compute a log determinant of an $n \times n$ positive definite matrix, and its derivatives - leading to prohibitive $\mathcal{O}(n^3)$ computations. We propose novel $\mathcal{O}(n)$ approaches to estimating these quantities from only fast matrix vector multiplications (MVMs). These stochastic approximations are based on Chebyshev, Lanczos, and surrogate models, and converge quickly even for kernel matrices that have challenging spectra. We leverage these approximations to develop a scalable Gaussian process approach to kernel learning. We find that Lanczos is generally superior to Chebyshev for kernel learning, and that a surrogate approach can be highly efficient and accurate with popular kernels.

Vlado Menkovski, Zharko Aleksovski, Axel Saalbach et al.

Convolutional neural networks demonstrated outstanding empirical results in computer vision and speech recognition tasks where labeled training data is abundant. In medical imaging, there is a huge variety of possible imaging modalities and contrasts, where annotated data is usually very scarce. We present two approaches to deal with this challenge. A network pretrained in a different domain with abundant data is used as a feature extractor, while a subsequent classifier is trained on a small target dataset; and a deep architecture trained with heavy augmentation and equipped with sophisticated regularization methods. We test the approaches on a corpus of X-ray images to design an anatomy detection system.

William Herlands, Andrew Wilson, Hannes Nickisch et al.

We present a scalable Gaussian process model for identifying and characterizing smooth multidimensional changepoints, and automatically learning changes in expressive covariance structure. We use Random Kitchen Sink features to flexibly define a change surface in combination with expressive spectral mixture kernels to capture the complex statistical structure. Finally, through the use of novel methods for additive non-separable kernels, we can scale the model to large datasets. We demonstrate the model on numerical and real world data, including a large spatio-temporal disease dataset where we identify previously unknown heterogeneous changes in space and time.

Andrew Gordon Wilson, Christoph Dann, Hannes Nickisch

We introduce a framework and early results for massively scalable Gaussian processes (MSGP), significantly extending the KISS-GP approach of Wilson and Nickisch (2015). The MSGP framework enables the use of Gaussian processes (GPs) on billions of datapoints, without requiring distributed inference, or severe assumptions. In particular, MSGP reduces the standard $O(n^3)$ complexity of GP learning and inference to $O(n)$, and the standard $O(n^2)$ complexity per test point prediction to $O(1)$. MSGP involves 1) decomposing covariance matrices as Kronecker products of Toeplitz matrices approximated by circulant matrices. This multi-level circulant approximation allows one to unify the orthogonal computational benefits of fast Kronecker and Toeplitz approaches, and is significantly faster than either approach in isolation; 2) local kernel interpolation and inducing points to allow for arbitrarily located data inputs, and $O(1)$ test time predictions; 3) exploiting block-Toeplitz Toeplitz-block structure (BTTB), which enables fast inference and learning when multidimensional Kronecker structure is not present; and 4) projections of the input space to flexibly model correlated inputs and high dimensional data. The ability to handle many ($m \approx n$) inducing points allows for near-exact accuracy and large scale kernel learning.

Andrew Gordon Wilson, Hannes Nickisch

We introduce a new structured kernel interpolation (SKI) framework, which generalises and unifies inducing point methods for scalable Gaussian processes (GPs). SKI methods produce kernel approximations for fast computations through kernel interpolation. The SKI framework clarifies how the quality of an inducing point approach depends on the number of inducing (aka interpolation) points, interpolation strategy, and GP covariance kernel. SKI also provides a mechanism to create new scalable kernel methods, through choosing different kernel interpolation strategies. Using SKI, with local cubic kernel interpolation, we introduce KISS-GP, which is 1) more scalable than inducing point alternatives, 2) naturally enables Kronecker and Toeplitz algebra for substantial additional gains in scalability, without requiring any grid data, and 3) can be used for fast and expressive kernel learning. KISS-GP costs O(n) time and storage for GP inference. We evaluate KISS-GP for kernel matrix approximation, kernel learning, and natural sound modelling.