ThesesPage

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

Diploma, Master and Bachelor Theses

Running Theses

Photorealistic Rendering of Training Data for Object Detection and Pose Estimation with a Physics Engine (DA/MA/BA)

3D Object Detection is essential for many tasks such a Robotic Manipulation or Augmented Reality. Nevertheless, recording appropriate real training data is difficult and time consuming. Due to this, many approaches rely on using synthetic data to train a Convolutional Neural Network. However, those approaches often suffer from overfitting to the synthetic world and do not generalize well to unseen real scenes. There are many works that try to address this problem. In this work we try to follow , and intend to render photorealistic scenes in order to cope with this domain gap. Therefore, we will use a physics engine to generate physically plausible poses and use ray-tracing to render high-quality scenes. In this particular work, we will extend another thesis to improve the renderings' quality as e.g. enhance the rendering realism in terms of lightning and reflection.
Supervisor:Fabian Manhardt, Johanna Wald
Director:Federico Tombari
Student:
start-end:21.02.2020 -
Deep Learning for Tool Detection and Tracking in Microsurgery (Bachelor Thesis)

The aim of this project is the investigation of the state-of-the-art deep learning architectures and frameworks with the purpose of detection and tracking of instruments in retinal microsurgeries. An implementation of a deep learning based instrument detection workflow shall be provided at the end of the project.
Supervisor:Hasan Sarhan, Dr. Mehmet Yigitsoy
Director:Prof. Nassir Navab
Student:Luca Alessandro Dombetzki
start-end:01.04.2018 -
Uncertainty Aware Methods for Camera Pose Estimation in Images and 3-Dimensional Data (Project)

Camera pose estimation is the term for determining the 6-DoF rotation and translation parameters of a camera. It is now a key technology in enabling multitudes of applications such as augmented reality, autonomous driving, human computer interaction and robot guidance. For decades, vision scholars have worked on finding the unique solution of this problem. Yet, this trend is witnessing a fundamental change. The recent school of thought has begun to admit that for our highly complex and ambiguous real environments, obtaining a single solution is not sufficient. This has led to a paradigm shift towards estimating rather a range of solutions in the form of full probability or at least explaining the uncertainty of camera pose estimates. Thanks to the advances in Artificial Intelligence, this important problem can now be tackled via machine learning algorithms that can discover rich and powerful representations for the data at hand. In collaboration, TU Munich and Stanford University plan to devise and implement generative methods that can explain uncertainty and ambiguity in pose predictions. In particular, our aim is to bridge the gap between 6DoF pose estimation either from 2D images/3D point sets and uncertainty quantification through multimodal variational deep methods.
Supervisor:Dr. Shadi Albarqouni, Dr. Tolga Birdal
Director:Prof. Dr. Nassir Navab, Prof. Dr. Leonidas Guibas
Student:Mai Bui, Haowen Deng
start-end:01.01.2020 -
3D Pedestrian Detection and Pose Estimation (Hiwi)

Autonomous driving systems are right on the corner and one key concern around the development and social acceptance of such systems is safeguarding. In this project, we want to look at the task of pedestrian detection from LiDAR? point clouds and their 3D pose estimation from the RGB camera input. 3D object detection from sparse point cloud data and multiple pedestrian 3D pose estimation are two challenging tasks and therefore active research fields in both academia and industry. In this project, we want to integrate the state of the art deep learning methods, train models on synthetic renderings and improve their performance based on safeguarding KPIs defined.
Supervisor:Mahdi Saleh
Director:Federico Tombari
Student:
start-end: -
Evaluation of real-time dense reconstruction for robotic navigation (Master Thesis)

Supervisor:Nikolas Brasch
Director:Federico Tombari
Student:
start-end: -
Distributed SLAM - Jointly mapping 3D Geometry (DA/MA/BA)

Exploring an unknown scene and self-positioning within it, is a common and well-studied problem in Computer Vision which is known as SLAM (Simultaneous Localization and Mapping). Core fields of application are autonomous cooperative robotics and vehicles as well as tracking and detection systems in the medical domain. Traditional methods target a single system, equipped with image sensors, exploring the scene and building up a map for localization (e.g. a single robot or drone moving within an unknown environment). New approaches also incorporate information from other sensors such as IMUs, gyro or GPS. Another objective for the determination of the position is outside-in-tracking of an object via marker tracking with external sensors, thus providing the relative position of an object with respect to the tracking system. To overcome the line-of-sight problem of outside-in-tracking, and the singularity constraint of traditional SLAM methods, the project aims to develop a distributed SLAM approach. Multiple systems (referred to as sensor nodes hereafter), equipped with an image sensor, contribute to a common map of the scene for localization, while being also tracked by outside-in-tracking for accuracy. Thus, accuracy and applicability can be elevated with a distributed SLAM approach, combining the information of multiple sensor nodes and an external tracking system. Furthermore, the necessity of complicated and error prone calibration processes for individual systems within one application scenario can be avoided. The objective is to develop a generative distributed SLAM approach for challenging scenes and applications. Features like loop detection and closing, pose graph optimization, re-localization and mapping should be extended to a distributed approach, also enabling scalability.
Supervisor:Patrick Ruhkamp, Benjamin Busam
Director:Prof. Dr. Nassir Navab
Student:Joe Bedard
start-end: -
Dynamic objects in dense reconstruction (DA/MA/BA)

Supervisor:Nikolas Brasch
Director:Federico Tombari
Student:
start-end: -
3D GAN for conditional medical image synthesis and cross-modality translation (Master Thesis)

GAN in 3D, especially in medical imaging application is challenging in many aspects, mainly due to the 'curse of dimensionality' and limited available data set. The goal of this project is to develop an optimum strategy to scale GAN in 3D that generalizes well for conditional medical image synthesis and cross-modality translation. The student will be provided with all-round support including good research environment, sufficient computational resources and active guidance to make the thesis successful.
Supervisor:Suprosanna Shit
Director:Prof. Bjoern Menze
Student:
start-end: -
3D Human Pose Estimation from RGB Images (DA/MA/BA)

Supervisor:Nikolas Brasch
Director:Federico Tombari
Student:
start-end: -
Incremental Learning For Robotic Grasping (Master Thesis)

Supervisor:Fabian Manhardt
Director:Nassir Navab
Student:Pengyuan Wang
start-end: -
Rethinking Deep Learning based Monocular Depth Prediction (Master Thesis)

We are looking for a motivated student who wants to work in the topic of monocular depth prediction using deep learning. Predicting depth from a single color image is a challenging and under-constrained task, and as such, active research is happening that incorporates CNNs. Recent works typically do not enforce an understanding for the objects in the scene. In contrast, our goal is to rethink depth prediction and create an object-aware model, which might lead to more accurate depth.
Supervisor:Helisa Dhamo
Director:Federico Tombari
Student:
start-end: -
Learning to learn: Which data we have to annotate first in medical applications? (Master Thesis)

Although the semi-supervised or unsupervised learning has been developed recently, the performance of them is still bound to the performance of fully-supervised learning. However, the cost of the annotation is extremely high in medical applications. It requires medical specialists (radiologists or pathologists) required to annotate the data. For those reasons, it is almost impossible to annotate all available dataset and sometimes, the only a subset of a dataset is possible to be selected for annotation due to the limited budget. Active learning is the research field which tries to deal with this problem [1-5]. Previous studies have been conducted in mainly three approaches: an uncertainty-based approach, a diversity-based approach, and expected model change [3]. These studies have been verified that active learning has the potential to reduce annotation cost. In this project, we aim to propose a novel active learning method which learns a simple uncertainty calculator to select more informative data to learn the current deep neural networks in medical applications.
Supervisor:Dr. Seong Tae Kim
Director:Prof. Dr. Nassir Navab
Student:Farrukh Mushtaq
start-end: -
Continual and incremental learning with less forgetting strategy (Master Thesis)

Recently, deep learning has great success in various applications such as image recognition, object detection, and medical applications, etc. However, in the real world deployment, the number of training data (sometimes the number of tasks) continues to grow, or the data cannot be given at once. In other words, a model needs to be trained over time with the increase of the data collection in a hospital (or multiple hospitals). A new type of lesion could be also defined by medical experts. Then, the pre-trained network needs to be further trained to diagnose these new types of lesions with increased data. ‘Class-incremental learning’ is a research area that aims at training the learned model to add new tasks while retaining the knowledge acquired in the past tasks. It is challenging because DNNs are easy to forget previous tasks when learning new tasks (i.e. catastrophic forgetting). In real-world scenarios, it is difficult to store all training data which was used when training DNN at the previous time due to the privacy issues of medical data. In this project, we will develop a solution to this problem in medical applications by investigating an effective and novel learning method.
Supervisor:Dr. Seong Tae Kim
Director:Prof. Dr. Nassir Navab
Student:Afshar Kakaei
start-end: -
Multiple sclerosis lesion segmentation from Longitudinal brain MRI (IDP)

Longitudinal medical data is defined that imaging data are obtained at more than one time-point where subjects are scanned repeatedly over time. Longitudinal medical image analysis is a very important topic because it can solve some difficulties which are limited when only spatial data is utilized. Temporal information could provide very useful cues for accurately and reliably analyzing medical images. To effectively analyze temporal changes, it is required to segment region-of-interest accurately in a short time. In the series of images acquired over multiple times of imaging, available cues for segmentation become richer with the intermediate predictions. In this project, we will investigate a way to fully exploit this rich source of information.
Supervisor:Dr. Seong Tae Kim, Ashkan Khakzar
Director:Prof. Dr. Nassir Navab
Student:Moiz Sajid, Stefan Denner
start-end: -
Understanding Medical Images to Generate Reliable Medical Report (Project)

The reading and interpretation of medical images are usually conducted by specialized medical experts [1]. For example, radiology images are read by radiologists and they write textual reports to describe the findings regarding each area of the body examined in the imaging study. However, writing medical-imaging reports requires experienced medical experts (e.g. experienced radiologists or pathologists) and it is time-consuming [2]. To assist in the administrative duties of writing medical-imaging reports, in recent years, a few research efforts have been devoted to investigating whether it is possible to automatically generate medical image reports for given medical image [3-8]. These methods are usually based on the encoder-decoder architecture which has been widely used for image captioning [9-10]. In this project, a novel automatic medical report generation method is investigated. It is challenging to generate accurate medical reports with large variation due to the high complexity in the natural language [11]. So, the traditional captioning methods suffer a problem where the model duplicates a completely identical sentence of the training set. To address the aforementioned limitations, this project focuses on the development of a reliable medical report generation method.
Supervisor:Dr. Shadi Albarqouni, Dr. Seong Tae Kim
Director:Prof. Dr. Nassir Navab
Student:Hossain Shaikh Saadi
start-end: -
Neural solver for PDEs (Hiwi)

Supervisor:Suprosanna Shit
Director:Bjoern Menze
Student:
start-end: -
Robust training of neural networks under noisy labels (Master Thesis)

The performance of supervised learning methods highly depends on the quality of the labels. However, accurately labeling a large number of datasets is a time-consuming task, which sometimes results in mismatched labeling. When the neural networks are trained with noisy data, it might be biased to the noisy data. Therefore the performance of the neural networks could be poor. While label noise has been widely studied in the machine learning society, only a few studies have been reported to identify or ignore them during the process of training. In this project, we will investigate the way to train the neural network under noisy data robustly. In particular, we will focus on exploring effective learning strategies and loss correction methods to address the problem.
Supervisor:Dr. Seong Tae Kim, Dr. Shadi Albarqouni
Director:Prof. Dr. Nassir Navab
Student:Cagri Yildiz
start-end: -
Self-supervised learning for out-of-distribution detection in medical applications (Master Thesis)

Although recent neural networks have achieved great successes when the training and the testing data are sampled from the same distribution, in real-world applications, it is unnatural to control the test data distribution. Therefore, it is important for neural networks to be aware of uncertainty when new kinds of inputs (which is called out-of-distribution) are given. In this project, we consider the problem of out-of-distribution detection in neural networks. In particular, we will develop a novel self-supervised learning approach for out-of-distribution detection in medical applications.
Supervisor:Dr. Seong Tae Kim
Director:Prof. Dr. Nassir Navab
Student:Abinav Ravi Venkatakrishnan
start-end: -
Development of spatio-temporal segmentation model for tumor volume calculation in micro-CT (Master Thesis)

To develop a spatio-temporal segmentation model where the network is exposed to previous temporal information and builds this complex mapping to segment a given mouse micro-CT image to allow accurate tumor volume calculations. A dataset with micro-CT scans of over 69 mice with repeat imaging is available with ground truth annotations. Mice were either treated with radiotherapy or left untreated. The small animal data act as a surrogate for clinical datasets treated with MR-linac technology, which requires automatic spatio-temporal segmentation.
Supervisor:Dr. Shadi Albarqouni, Dr. Seong Tae Kim, Dr. Guillaume Landry
Director:Prof. Dr. Nassir Navab
Student:Tetiana Klymenko
start-end: -
Trajectory Validation using Deep Learning Methods (Master Thesis)

Supervisor:Nikolas Brasch
Director:Federico Tombari
Student:
start-end: -

Finished Theses



Edit | Attach | Refresh | Diffs | More | Revision r1.14 - 02 Jan 2019 - 11:39 - TobiasLasser

Lehrstuhl für Computer Aided Medical Procedures & Augmented Reality    rss.gif