Tobias Würfl

Deep Learning Computed Tomography

In this paper, we demonstrate that image reconstruction can be expressed in terms of neural networks. We show that filtered back-projection can be mapped identically onto a deep neural network architecture. As for the case of iterative reconstruction, the straight forward realization as matrix multiplication is not feasible. Thus, we propose to compute the back-projection layer efficiently as fixed function and its gradient as projection operation. This allows a data-driven approach for joint optimization of correction steps in projection domain and image domain. As a proof of concept, we demonstrate that we are able to learn weightings and additional filter layers that consistently reduce the reconstruction error of a limited angle reconstruction by a factor of two while keeping the same computational complexity as filtered back-projection. We believe that this kind of learning approach can be extended to any common CT artifact compensation heuristic and will outperform hand-crafted artifact correction methods in the future.

A Deep Learning Architecture for Limited-Angle Computed Tomography Reconstruction

Limited-angle computed tomography suffers from missing data in the projection domain, which results in intensity inhomogeneities and streaking artifacts in the image domain. We address both challenges by a two-step deep learning architecture: First, we learn compensation weights that account for the missing data in the projection domain and correct for intensity changes. Second, we formulate an image restoration problem as a variational network to eliminate coherent streaking artifacts. We perform our experiments on realistic data and we achieve superior results for destreaking compared to state-of-the-art non-linear filtering methods in literature. We show that our approach eliminates the need for manual tuning and enables joint optimization of both correction schemes.

Some Investigations on Robustness of Deep Learning in Limited Angle Tomography

In computed tomography, image reconstruction from an insufficient angular range of projection data is called limited angle tomography. Due to missing data, reconstructed images suffer from artifacts, which cause boundary distortion, edge blurring, and intensity biases. Recently, deep learning methods have been applied very successfully to this problem in simulation studies. However, the robustness of neural networks for clinical applications is still a concern. It is reported that most neural networks are vulnerable to adversarial examples. In this paper, we aim to investigate whether some perturbations or noise will mislead a neural network to fail to detect an existing lesion. Our experiments demonstrate that the trained neural network, specifically the U-Net, is sensitive to Poisson noise. While the observed images appear artifact-free, anatomical structures may be located at wrong positions, e.g. the skin shifted by up to 1 cm. This kind of behavior can be reduced by retraining on data with simulated Poisson noise. However, we demonstrate that the retrained U-Net model is still susceptible to adversarial examples. We conclude the paper with suggestions towards robust deep-learning-based reconstruction.

Adversarial and Perceptual Refinement for Compressed Sensing MRI Reconstruction.

Deep learning approaches have shown promising performance for compressed sensing-based Magnetic Resonance Imaging. While deep neural networks trained with mean squared error (MSE) loss functions can achieve high peak signal to noise ratio, the reconstructed images are often blurry and lack sharp details, especially for higher undersampling rates. Recently, adversarial and perceptual loss functions have been shown to achieve more visually appealing results. However, it remains an open question how to (1) optimally combine these loss functions with the MSE loss function and (2) evaluate such a perceptual enhancement. In this work, we propose a hybrid method, in which a visual refinement component is learnt on top of an MSE loss-based reconstruction network. In addition, we introduce a semantic interpretability score, measuring the visibility of the region of interest in both ground truth and reconstructed images, which allows us to objectively quantify the usefulness of the image quality for image post-processing and analysis. Applied on a large cardiac MRI dataset simulated with 8-fold undersampling, we demonstrate significant improvements (\(p<0.01\)) over the state-of-the-art in both a human observer study and the semantic interpretability score.

Optimizing R with SparkR on a commodity cluster for biomedical research

BACKGROUND AND OBJECTIVES Medical researchers are challenged today by the enormous amount of data collected in healthcare. Analysis methods such as genome-wide association studies (GWAS) are often computationally intensive and thus require enormous resources to be performed in a reasonable amount of time. While dedicated clusters and public clouds may deliver the desired performance, their use requires upfront financial efforts or anonymous data, which is often not possible for preliminary or occasional tasks. We explored the possibilities to build a private, flexible cluster for processing scripts in R based on commodity, non-dedicated hardware of our department. METHODS For this, a GWAS-calculation in R on a single desktop computer, a Message Passing Interface (MPI)-cluster, and a SparkR-cluster were compared with regards to the performance, scalability, quality, and simplicity. RESULTS The original script had a projected runtime of three years on a single desktop computer. Optimizing the script in R already yielded a significant reduction in computing time (2 weeks). By using R-MPI and SparkR, we were able to parallelize the computation and reduce the time to less than three hours (2.6 h) on already available, standard office computers. While MPI is a proven approach in high-performance clusters, it requires rather static, dedicated nodes. SparkR and its Hadoop siblings allow for a dynamic, elastic environment with automated failure handling. SparkR also scales better with the number of nodes in the cluster than MPI due to optimized data communication. CONCLUSION R is a popular environment for clinical data analysis. The new SparkR solution offers elastic resources and allows supporting big data analysis using R even on non-dedicated resources with minimal change to the original code. To unleash the full potential, additional efforts should be invested to customize and improve the algorithms, especially with regards to data distribution.

Patch-based Carcinoma Detection on Confocal Laser Endomicroscopy Images - A Cross-site Robustness Assessment

Deep learning technologies such as convolutional neural networks (CNN) provide powerful methods for image recognition and have recently been employed in the field of automated carcinoma detection in confocal laser endomicroscopy (CLE) images. CLE is a (sub-)surface microscopic imaging technique that reaches magnifications of up to 1000x and is thus suitable for in vivo structural tissue analysis. In this work, we aim to evaluate the prospects of a priorly developed deep learning-based algorithm targeted at the identification of oral squamous cell carcinoma with regard to its generalization to further anatomic locations of squamous cell carcinomas in the area of head and neck. We applied the algorithm on images acquired from the vocal fold area of five patients with histologically verified squamous cell carcinoma and presumably healthy control images of the clinically normal contra-lateral vocal cord. We find that the network trained on the oral cavity data reaches an accuracy of 89.45% and an area-under-the-curve (AUC) value of 0.955, when applied on the vocal cords data. Compared to the state of the art, we achieve very similar results, yet with an algorithm that was trained on a completely disjunct data set. Concatenating both data sets yielded further improvements in cross-validation with an accuracy of 90.81% and AUC of 0.970. In this study, for the first time to our knowledge, a deep learning mechanism for the identification of oral carcinomas using CLE Images could be applied to other disciplines in the area of head and neck. This study shows the prospect of the algorithmic approach to generalize well on other malignant entities of the head and neck, regardless of the anatomical location and furthermore in an examiner-independent manner.

Towards Arbitrary Noise Augmentation—Deep Learning for Sampling from Arbitrary Probability Distributions

This paper proposes a fully connected neural network model to map samples from a uniform distribution to samples of any explicitly known probability density function. During the training, the Jensen-Shannon divergence between the distribution of the models output and the target distribution is minimized. We experimentally demonstrate that our model converges towards the desired state. It provides an alternative to existing sampling methods such as inversion sampling, rejection sampling, Gaussian mixture models and Markov-Chain-Monte-Carlo. Our model has high sampling efficiency and is easily applied to any probability distribution, without the need of further analytical or numerical calculations. It can produce correlated samples, such that the output distribution converges faster towards the target than for independent samples. But it is also able to produce independent samples, if single values are fed into the network and the input values are independent as well. We focus on one-dimensional sampling, but additionally illustrate a two-dimensional example with a target distribution of dependent variables.

Multiple Device Segmentation for Fluoroscopic Imaging Using Multi-task Learning

For endovascular aortic repair (EVAR), integrating preoperative information of the aortic anatomy with intraoperative fluoroscopy can aid in reducing radiation exposure, contrast agent and procedure time. However, the quality of this fusion may deteriorate over the course of the intervention due to patient movement or deformation of the vasculature caused by interventional tools. Automatically detecting the instruments present in the X-ray image can help to assess the degree of deterioration, trigger automatic re-registration or aid in automatic workflow phase detection and process modeling. In this work, we investigate a flexible approach to segment different devices based on fully convolutional neural networks using multi-task learning. We evaluate the proposed approach on a set of 38 X-ray images acquired during EVAR interventions by targeting the segmentation of aortic stents, stiff guidewires and pigtail catheters. We compare the results to the performance of single-task networks. We manage to keep similar performance compared to single-task networks with Dice coefficients between 0.95 and 0.80 depending on the device, while speeding up computation by a factor of two.

Deep Learning Fundamentals

Deep learning has received a lot of attention in the machine learning community. Successful applications from speech recognition or computer vision are already part of our daily life. Much effort has been devoted to transferring this success to medical image computing.