Index
Geometric Domain Adaptation for CBCT Segmentation
In this project, we perform a computational domain transfer to introduce cone-beam artifacts to the training data. We evaluate its impact on the results of supervised training for the segmentation of the lungs. For this, already labeled CT volumes are reconstructed to artificial CBCT volumes without a complex deep learning-based method, like introduced by Jia X et al.,5 but rather by computational reconstruction. The purpose is to have a network for stable segmentation on real CBCT volumes. A major advantage of our approach is that the artificial
CBCT volumes can not only be computed easily from thoracic CT volumes but also the pixel-wise segmentation can be re-used without putting in the great effort of labeling. This allows for supervised training.
Realistic Simulation of Collimated X-Ray images for Collimator Edge Segmentation using Deep Learning
Collimator detection in X-ray systems has long posed a challenge, particularly when information about the detector’s position relative to the source is either unreliable or completely unavailable. In this paper [1], we introduce a physically motivated image processing pipeline designed to simulate the intricate characteristics of collimator shadows in X-ray images. The primary objective of this pipeline is to address the scarcity of training data for deep neural networks, which are increasingly promising for collimator detection. By applying the pipeline to deep networks initially limited by small datasets, our approach equips them with the necessary information to learn and generalize effectively.
Our pipeline is a comprehensive solution that leverages several key components to
generate realistic collimator images. Employing randomized labels to describe collimator shapes and their respective locations ensures diversity and representativeness. In addition, we integrate a convolution kernel based scattered radiation simulation mechanism, which is a crucial factor in real-world X-ray imaging. To complete the simulation process, we introduce Poisson noise to replicate the inherent characteristics of collimator shadows in X-ray images.
Comparing the simulated data with real collimator shadows demonstrates the authenticity of our approach and its potential to bridge the gap between synthetic and real-world data. Moreover, incorporating simulated data into our deep learning framework not only serves as a valid substitute for real collimators but also significantly improves generalization in real-world applications, holding great promise for the field of collimator detection.
This work was presented at the DALI workshop at the MICCAI conference in Vancouver, Canada and was published in the proceedings:
1. El-Zein B, Eckert D, Weber T, Rohleder M, Ritschl L, Kappler S et al. A Realistic Collimated
X-Ray Image Simulation Pipeline. Data Augmentation, Labelling, and Imperfections – Third
MICCAIWorkshop, DALI 2023, Held in Conjunction with MICCAI 2023, Vancouver, October
12, 2023, Proceedings. Springer Nature. 2023, (in press)
Unsupervised Domain Adaptation Using Contrastive Learning for Multi-modal Cardiac MR Segmentation
Deep Learning Computed Tomography based on the Defrise and Clack Algorithm for Specific CBCT Orbits
The RoboCT system enables the exploration of questions that are not feasible with traditional rotating table or gantry setups. By placing the source and detector freely around the object, non-standard trajectories (e.g., circle, helix) can be obtained, which is essential for various reasons such as complex objects, ROI reconstructions, limited angle imaging, etc. However, reconstructing data acquired with these trajectories using an FBP-based algorithm like FDK is not straightforward. Currently, deviations from the circular trajectory are reconstructed using an algebraic reconstruction technique (ART). The parameterization of ART significantly affects the reconstruction quality, especially under challenging conditions such as limited angle, sparse sampling, or truncation artefacts. ART also has drawbacks, including longer computation time compared to FBP and the need to start the iterative process after data acquisition.
Theoretical descriptions exist for FBP-based reconstruction for general trajectories [1, 2]. However, the filtering step cannot be performed with shift-invariant filter kernels like in FDK. Instead, trajectory-specific filter kernels must be derived and determined, making the process complex and time-consuming.
This master’s thesis aims to investigate the possibility of learning invariant or variant filters specific to trajectories using Known Operator Learning [3]. Building on previous work in filter learning [4,5], the implementation will be carried out using the Pyronn framework [6]. The study will also explore whether these filters can be learned purely from simulated data using a specially created phantom and their generalization to real data with different objects under the specified trajectory, based on previous research [5].
[1] Defrise, Michel, and Rolf Clack. “A cone-beam reconstruction algorithm using shift-variant filtering and cone-beam backprojection.” IEEE transactions on medical imaging 13.1 (1994): 186-195.
[2] Oeckl, Steven. “Rekonstruktionsverfahren mit der approximativen Inversen und einer neuen Formel zur Inversion der Röntgen-Transformation.” (2014).
[3] Maier, Andreas K., et al. “Learning with known operators reduces maximum error bounds.” Nature machine intelligence 1.8 (2019): 373-380.
[4] Syben, Christopher, et al. “Precision learning: reconstruction filter kernel discretization.” arXiv preprint arXiv:1710.06287 (2017).
[5] Syben, Christopher, et al. “Known operator learning enables constrained projection geometry conversion: Parallel to cone-beam for hybrid MR/X-ray imaging.” IEEE Transactions on Medical Imaging 39.11 (2020): 3488-3498.
[6] Syben, Christopher, et al. “PYRO-NN: Python reconstruction operators in neural networks.” Medical physics 46.11 (2019): 5110-5115.
Robot Movement Planning for Obstacle Avoidance using Reinforcement Learning
Obstacle avoidance for robotic arms is an important issue in robot control. Limited by factors such as equipment, cost, and labor, some application scenarios require the robot to have the ability to plan its own movement to reach the goal position or state. 
In real working environments, where obstacles’ properties are various, the traditional search algorithms are difficult to adapt to large-scale space and continuous action requirements. Artificial potential field (APF) method is a widely used obstacle avoidance path planning method, but it alsohas some shortcomings and will fall into local optimality in some special situations. Based on APF method, reinforcement learning (RL) can theoretically achieve optimization in continuous space. Combined with some modifications to traditional APF method, we define states, and actions and design the reward function, which is regarded as an important part of reinforcement learning, to form a motion planning agent in the
3D world, so that the robot end-effector is able to reach the goal position, emphasizing the avoidance of collision between obstacles and the whole robot arm.
Sinogram Analysis Using Attention U-Net: A Methodological Approach to Defect Detection and Localization in Parallel Beam Computed Tomography
The emergence of deep learning has ushered in a transformative era within the realm of image processing, notably in the context of Computed Tomography (CT). Nevertheless, it is noteworthy that a majority of image processing algorithms traditionally rely on processed or reconstructed images, often overlooking the raw sensor data. This thesis, however, shifts its focus toward the utilization of unprocessed computed tomography data, which we refer to as sinogram. Within this framework, we present a comprehensive three-step deep learning algorithm, leveraging a UNet-based architecture, designed to identify and analyze defects within objects without resorting to image reconstruction. The initial phase entails sinogram segmentation, facilitating the extraction of defect masks within the sinogram. Subsequently, instance segmentation is employed to effectively segregate these masks, resulting in their individualization. Lastly, the isolated masks are subjected to thorough defect analysis. Our research endeavors encompass comprehensive experimentation, conducted on both simulated datasets and real-world data.