Simon Pezold

Multi-dimensional Gated Recurrent Units for the Segmentation of Biomedical 3D-Data

We present a supervised deep learning method to automatically segment 3D volumes of biomedical image data. The presented method takes advantage of a neural network with the main layers consisting of multi-dimensional gated recurrent units. We apply an on-the-fly data augmentation technique which allows for accurate estimations without the need for either a huge amount of training data or advanced data pre- or postprocessing. We show that our method performs amongst the leading techniques on a popular brain segmentation challenge dataset in terms of speed, accuracy and memory efficiency. We describe in detail advantages over a similar method which uses the well-established long short-term memory.

Automatic Segmentation of the Spinal Cord Using Continuous Max Flow with Cross-sectional Similarity Prior and Tubularity Features

Segmenting tubular structures from medical image data is a common problem; be it vessels, airways, or nervous tissue like the spinal cord. Many application-specific segmentation techniques have been proposed in the literature, but only few of them are fully automatic and even fewer approaches maintain a convex formulation. In this paper, we show how to integrate a cross-sectional similarity prior into the convex continuous max-flow framework that helps to guide segmentations in image regions suffering from noise or artefacts. Furthermore, we propose a scheme to explicitly include tubularity features in the segmentation process for increased robustness and measurement repeatability. We demonstrate the performance of our approach by automatically segmenting the cervical spinal cord in magnetic resonance images, by reconstructing its surface, and acquiring volume measurements.

With Gaze Tracking Toward Noninvasive Eye Cancer Treatment

We present a new gaze tracking-based navigation scheme for proton beam radiation of intraocular tumors and we show the technical integration into the treatment facility. Currently, to treat a patient with such a tumor, a medical physicist positions the patient and the affected eye ball such that the radiation beam targets the tumor. This iterative eye positioning mechanism requires multiple X-rays, and radio-opaque clips previously sutured on the target eyeball. We investigate a possibility to replace this procedure with a noninvasive approach using a 3-D model-based gaze tracker. Previous work does not cover a comparably extensive integration of a gaze tracking device into a state-of-the-art proton beam facility without using additional hardware, such as a stereo optical tracking system. The integration is difficult because of limited available physical space, but only this enables to quantify the overall accuracy. We built a compact gaze tracker and integrated it into the proton beam radiation facility of the Paul Scherrer Institute in Villigen, Switzerland. Our results show that we can accurately estimate a healthy volunteers point of gaze, which is the basis for the determination of the desired initial eye position. The proposed method is the first crucial step in order to make the proton therapy of the eye completely noninvasive.

Automatic, Robust, and Globally Optimal Segmentation of Tubular Structures

We present an automatic three-dimensional segmentation approach based on continuous max flow that targets tubular structures in medical images. Our method uses second-order derivative information provided by Frangi et al.’s vesselness feature and exploits it twofold: First, the vesselness response itself is used for localizing the tubular structure of interest. Second, the eigenvectors of the Hessian eigendecomposition guide our anisotropic total variation–regularized segmentation. In a simulation experiment, we demonstrate the superiority of anisotropic as compared to isotropic total variation–regularized segmentation in the presence of noise. In an experiment with magnetic resonance images of the human cervical spinal cord, we compare our automated segmentations to those of two human observers. Finally, a comparison with a dedicated state-of-the-art spinal cord segmentation framework shows that we achieve comparable to superior segmentation quality.

Augmented reality assisted laparoscopic partial nephrectomy

Computer assisted navigation is a widely adopted technique in neurosurgery and orthopedics. However, it is rarely used for surgeries on abdominal organs. In this paper, we propose a novel, noninvasive method based on electromagnetic tracking to determine the pose of the kidney. As a clinical use case, we show a complete surgical navigation system for augmented reality assisted laparoscopic partial nephrectomy. Experiments were performed ex vivo on pig kidneys and the evaluation showed an excellent augmented reality alignment error of 2.1 mm ± 1.2 mm.

A Semi-automatic Method for the Quantification of Spinal Cord Atrophy

Due to its high flexibility, the spinal cord is a particularly challenging part of the central nervous system for the quantification of nervous tissue changes. In this paper, a novel semi-automatic method is presented that reconstructs the cord surface from MR images and reformats it to slices that lie perpendicular to its centerline. In this way, meaningful comparisons of cord cross-sectional areas are possible. Furthermore, the method enables to quantify the complete upper cervical cord volume. Our approach combines graph cut for presegmentation, edge detection in intensity profiles for segmentation refinement, and the application of starbursts for reformatting the cord surface. Only a minimum amount of user input and interaction time is required. To quantify the limits and to demonstrate the robustness of our approach, its accuracy is validated in a phantom study and its precision is shown in a volunteer scan–rescan study. The method’s reproducibility is compared to similar published quantification approaches. The application to clinical patient data is presented by comparing the cord cross-sections of a group of multiple sclerosis patients with those of a matched control group, and by correlating the upper cervical cord volumes of a large MS patient cohort with the patients’ disability status. Finally, we demonstrate that the geometric distortion correction of the MR scanner is crucial when quantitatively evaluating spinal cord atrophy.

Adaptive Graph Diffusion Regularisation for Discontinuity Preserving Image Registration.

Registration of thoracic images is central when studying for example physiological changes of the lung. Due to sliding organ motion and intensity changes based on respiration the registration of thoracic images is challenging. We present a novel regularisation method based on adaptive anisotropic graph diffusion. Without the need of a mask it preserves discontinuities of the transformation at sliding organ boundaries and enforces smoothness in areas with similar motion. The graph diffusion regularisation provides a direct way to achieve anisotropic diffusion at sliding organ boundaries by reducing the weight of corresponding edges in the graph which cross the sliding interfaces. Since the graph diffusion is defined by the edge weights of the graph, we develop an adaptive edge weight function to detect sliding boundaries. We implement the adaptive graph diffusion regularisation method in the Demons registration framework. The presented method is tested on synthetic 2D images and on the public 4D-CT DIR-Lab data set, where we are able to correctly detect the sliding organ boundaries.

Spinal cord volume loss: A marker of disease progression in multiple sclerosis

Objective Cross-sectional studies have shown that spinal cord volume (SCV) loss is related to disease severity in multiple sclerosis (MS). However, long-term data are lacking. Our aim was to evaluate SCV loss as a biomarker of disease progression in comparison to other MRI measurements in a large cohort of patients with relapse-onset MS with 6-year follow-up. Methods The upper cervical SCV, the total brain volume, and the brain T2 lesion volume were measured annually in 231 patients with MS (180 relapsing-remitting [RRMS] and 51 secondary progressive [SPMS]) over 6 years on 3-dimensional, T1-weighted, magnetization-prepared rapid-acquisition gradient echo images. Expanded Disability Status Scale (EDSS) score and relapses were recorded at every follow-up. Results Patients with SPMS had lower baseline SCV (p < 0.01) but no accelerated SCV loss compared to those with RRMS. Clinical relapses were found to predict SCV loss over time (p < 0.05) in RRMS. Furthermore, SCV loss, but not total brain volume and T2 lesion volume, was a strong predictor of EDSS score worsening over time (p < 0.05). The mean annual rate of SCV loss was the strongest MRI predictor for the mean annual EDSS score change of both RRMS and SPMS separately, while correlating stronger in SPMS. Every 1% increase of the annual SCV loss rate was associated with an extra 28% risk increase of disease progression in the following year in both groups. Conclusion SCV loss over time relates to the number of clinical relapses in RRMS, but overall does not differ between RRMS and SPMS. SCV proved to be a strong predictor of physical disability and disease progression, indicating that SCV may be a suitable marker for monitoring disease activity and severity.

Automated Segmentation of Multiple Sclerosis Lesions Using Multi-dimensional Gated Recurrent Units

We analyze the performance of multi-dimensional gated recurrent units on automated lesion segmentation in multiple sclerosis. The segmentation of these pathologic structures is not trivial, since location, shape and size can be arbitrary. Furthermore, the inherent class imbalance of about 1 lesion voxel to 10 000 healthy voxels further exacerbates the correct segmentation. We introduce a new MD-GRU setup, using established techniques from the deep learning community as well as our own adaptations. We evaluate these modifications by comparing them to a standard MD-GRU network. We demonstrate that using data augmentation, selective sampling, residual learning and/or DropConnect on the RNN state can produce better segmentation results. Reaching rank #1 in the ISBI 2015 longitudinal multiple sclerosis lesion segmentation challenge, we show that a setup which combines these techniques can outperform the state of the art in automated lesion segmentation.

Direct Calibration of a Laser Ablation System in the Projective Voltage Space

Laser ablation is a widely adopted technique in many contemporary medical applications. However, it is new to use a laser to cut bone and perform general osteotomy surgical tasks with it. In this paper, we propose to apply the direct linear transformation algorithm to calibrate and integrate a laser deflecting tilting mirror into the affine transformation chain of a sophisticated surgical navigation system, involving next generation robots and optical tracking. Experiments were performed on synthetic input and real data. The evaluation showed a target registration error of 0.3mm ± 0.2mm in a working distance of 150 mm.