Page tree
Skip to end of metadata
Go to start of metadata


The seminar takes place continuously (also during the semester break). In order to arrange a talk, please register through our Oberseminar registration formThis can only be done by the project supervisors.


Mobile view: If you are having trouble seeing the schedule on your phone screen, please switch to the desktop version using the ≡ button on the top left corner.



PresentationDiscussion
Kick-offs10 min5 min
Finals20 min10 min


Schedule:




Date

Time

Student

Title

Type

Supervisor

30.Aug (Fr) 10:00 Jyotirmay Senapati Bayesian Deep Learning for Medical Image Segmentation MA Kick-Off Abhijit Guha Roy
30.Aug (Fr) 10:30 Mina Sedra Rigid 3D image stitching for thyroid MA Kick-Off Matthias Keicher
30.Aug (Fr) 11:00 Mayar Kamali Modeling uncertainty using Monte-Carlo Dropout for Proximal Femur Fractures BA Final Shadi Albarqouni
30.Aug (Fr) 11:30 Ricardo Smits GANFormer: Exploring GANs for Ultrasound Reconstruction MA Final Walter Simson
30.Aug (Fr) 12:00 Abhijeet Parida Few Shot Segmentation in Medical Imaging MA Final Shadi Albarqouni
30.Aug (Fr) 12:30 Berk Olcay DeepVOG - Oculomotor event detection in video-oculography MA Kick-Off Virginia Flanagin
30.Aug (Fr) 13:00 Aashish Trivedi Occlusion-aware 6D object pose estimation MA Kick-Off Fabian Manhardt
06.Sep (Fr) 10:30 HyunJun Jung Light-weight 6D Pose Tracking with Reinforcement Learning MA Kick-Off Benjamin Busam
06.Sep (Fr) 11:00 Christina Bukas Estimation of Vertebral Bodies prior to Fracture Damage using Generative Models MA Kick-Off Thomas Wendler
06.Sep (Fr) 11:30 Sebastian Lutz Automatic Segmentation of MR Spinal Images Using Knowledge Transfer from CT MA Kick-Off Thomas Wendler
06.Sep (Fr) 12:00 Peter Mortimer 3D Instances from a single RGB Image MA Final Federico Tombari
06.Sep (Fr) 12:30 Viktoria Markova Detection of Adversarial Attacks Using Feature Visualisation BA Final Manuel Nickel
06.Sep (Fr) 13:00 Taha Emre A Deep Learning Approach for Monitoring the Critical View of Safety in cholecystectomy Procedures MA Final Prof. Dr. Nassir Navab
06.Sep (Fr) 13:30 Marcel Schmitt Semantic Object-Level SLAM in Outdoor Environments MA Final Markus Herb
13.Sep (Fr) 11:30 Granit Bimbashi Dynamic Volume Estimation of Bone Cement during Vertebroplasty for Therapy Optimization MA Kick-Off Thomas Wendler
27.Sep (Fr) 10:30 Robert Graf Exploring CycleGANs for Anomaly Detection in Brain MRI BA Final Christoph Baur
04.Okt (Fr) 10:30 Stefan Walke Automatic CT Classification and Segmentation of Bone Cement after Osteoplasty for Therapy Optimization MA Kick-Off Thomas Wendler


Detailed information about the above presentations:



Date & Time

30.August 2019 -- 10:00

Title

Bayesian Deep Learning for Medical Image Segmentation

Student

Jyotirmay Senapati

Type

MA Kick-Off

Supervisor

Abhijit Guha Roy

Additional supervisors

Prof. Dr. Christian Wachinger

Director

Prof. Nassir Navab

Abstract

Bayesian Deep Learning has gained a lot of popularity in the Deep Learning community due to its ability to generate well calibrated outputs by associating a confidence score along with the prediction. This has high implications in using deep learning for safety critical applications like Medical diagnosis, Autonomous driving etc. Recently, alot of work has explored and proposed different strategies for variational inference of Deep Networks to estimate the confidence. It is very difficult to say which is better for what application. Towards this end we explore these different strategies to identify their effectiveness targeting the application of segmentation Quality Control.

Date & Time

30.August 2019 -- 10:30

Title

Rigid 3D image stitching for thyroid

Student

Mina Sedra

Type

MA Kick-Off

Supervisor

Matthias Keicher

Additional supervisors

Robert Bauer

Director

Prof. Nassir Navab

Abstract

Comparing the left and the right thyroid volumes and shapes is a crucial step towards thyroid disease diagnosis and therapy. While thyroids are usually scanned by ultrasound systems, 3D volumes of the left and the right thyroid can be acquired for better visualization and better diagnosis by using systems developed by PIUR GmbH. Although recent research has been excessively done to achieve a single view scan through mosaicing, the process seems to be imperfect and rather hard due to the difficulty of maximizing the matching pixels/boxes while minimizing the overlapping area between the 2 images at the same time, also the difficulty of finding the adequate initialization steps and needed transformation of the 2 volumes because of the shortness of the capture range of state of the art methods. To solve this problem, We propose a new mono-modal rigid transformation method with a high capture range for the initialization of volumes with small overlap for deformable registration.

Date & Time

30.August 2019 -- 11:00

Title

Modeling uncertainty using Monte-Carlo Dropout for Proximal Femur Fractures

Student

Mayar Kamali

Type

BA Final

Supervisor

Shadi Albarqouni

Director

Prof. Nassir Navab

Abstract

Fracture of the human fracture specially the long bones are among the most common reasons why patients visit the emergency rooms. X Rays are evaluated then by either expert radiologists or trauma surgeons for the classification of the fracture. Accurate fracture classification requires years of practice and experience as reflected by the reported inter-observer variability which is as low as 66% for residents and 71% for experts. The aim of our project is to compute the epistemic uncertainty of a neural network and model the uncertainty using the recently proposed Monte-Carlo Dropout to report the uncertainty in predictive classes. Epistemic uncertainty is an important quantity for the deployment of deep neural networks in safety-critical applications, since it represents how much one can trust predictions on new data.

Date & Time

30.August 2019 -- 11:30

Title

GANFormer: Exploring GANs for Ultrasound Reconstruction

Student

Ricardo Smits

Type

MA Final

Supervisor

Walter Simson

Additional supervisors

Magda Paschali

Director

Prof. Nassir Navab

Abstract

In this work, we present a novel approach utilizing a GAN-based architecture to learn ultrasound speckle filter from noisy ground truth data. The construction of our model enables an end-to-end framework for the direct reconstruction of ultrasound B-Mode images that contain high information content, while selectively filtering our speckle noise.

Date & Time

30.August 2019 -- 12:00

Title

Few Shot Segmentation in Medical Imaging

Student

Abhijeet Parida

Type

MA Final

Supervisor

Shadi Albarqouni

Additional supervisors

Ari Tran

Director

Prof. Nassir Navab

Abstract

Semantic segmentation does a pixel-wise classification to assign a class or background to each pixel of an image. This problem requires a very large data set of pixel-level annotations, which is often unavailable or very costly to create. The aim of this project is to build a state of the art low shot deep learning technique for medical images, which can from a few dense or sparse annotated medical image labels derive the semantic segmentation of a new previously unseen class.

Date & Time

30.August 2019 -- 12:30

Title

DeepVOG - Oculomotor event detection in video-oculography

Student

Berk Olcay

Type

MA Kick-Off

Supervisor

Virginia Flanagin

Additional supervisors

Seyed-Ahmad Ahmadi

Director

Prof. Nassir Navab

Abstract

This M.Sc. thesis is about analysing eye movements recorded during video-oculography examinations. The aim is a detection of oculomotor events using deep learning, which can be an aid to diagnose physiological and psychological disorders. For example, nystagmus classification is important for cerebellar disorders, while fixation and saccade events are important for analyzing attention disorders. Normally, these analyses are done with different semi-automatic algorithms that require a patient-specific and time-consuming adaptation of detection parameters. The aim is to automize it using deep learning. The network training will be done with oculomotor data acquired at our clinic and publicly available and validated fixation/saccade data. Currently, we consider using a temporal CNN architecture to accomplish this task, which is causal and considers the last 30 seconds of eye movements in order to predict multi-label classification for every time step.

Date & Time

30.August 2019 -- 13:00

Title

Occlusion-aware 6D object pose estimation

Student

Aashish Trivedi

Type

MA Kick-Off

Supervisor

Fabian Manhardt

Additional supervisors

Helisa Dhamo

Director

PD Dr. Federico Tombari

Abstract

Pose estimation of objects using monocular vision is an active area of research. It works quite well under the condition that the objects in the scene are completely visible. But when the objects are occluded, pose estimation becomes unreliable. Overcoming occlusion is another active area of research in computer vision where some of the work has been done on inpainting the objects that are occluded. The thesis will combine these two approaches by inpainting the occluded parts of objects in the scene and then estimating pose of those objects more reliably.

Date & Time

06.September 2019 -- 10:30

Title

Light-weight 6D Pose Tracking with Reinforcement Learning

Student

HyunJun Jung

Type

MA Kick-Off

Supervisor

Benjamin Busam

Director

Prof. Nassir Navab

Abstract

6D pose tracking is an important problem in various applications such as augmented reality and robotics. Although various deep learning based 6D pose tracking methods have been already developed and give reliable results, they are mostly computationally expensive and need training for each individual object. This thesis explores a new way of tracking 6D poses for objects in RGB images leveraging learning based classification together with reinforcement learning. The goal is a computationally tracktable pipeline (ideally several convolutional layers and fully connected layers at most). Starting from a rough initial pose of a moving object, an online frame-to-frame tracker adjusts small incremental discrete poses to the observation with an extra input given by a real-time renderer that provides misalignment information between the pose of the object in the RGB image and the current pose estimate. The network has two training phases: A first classifier for tracking actions at given misalignments is trained fully supervised. Then, the network will be finetuned with reinforcement learning exploiting sequential video information. The results are planned to be evaluated on common datasets as well as with a prototype application.

Date & Time

06.September 2019 -- 11:00

Title

Estimation of Vertebral Bodies prior to Fracture Damage using Generative Models

Student

Christina Bukas

Type

MA Kick-Off

Supervisor

Thomas Wendler

Additional supervisors

Magda Paschali

Director

Prof. Nassir Navab

Abstract

A vertebral fracture is a serious injury that causes severe pain to the patient and can occur as a result of an accident, but also due to diseases and illnesses, mainly osteoporosis. In such cases minimally invasive surgery is often performed, namely vertebroplasty and kyphoplasty, to restore vertebral body height and alleviate the patient from the pain. These methods involve injection of a cement-like material directly into the vertebral body of the fractured vertebra. However, since the original volume of the fractured vertebral body is unknown the amount of cement injected is mainly based on experience and common practices. We propose a neural network which can reconstruct the original shape of the healthy vertebra based on information provided by its neighboring vertebrae, available to us through the post-fracture CT scan of the patient. Our aim is to make the clinician aware of the volume of the vertebra prior to the fracture and therefore also aware of the exact cement amount needed during surgery, in order to bring the vertebral body as close to its original form as possible.

Date & Time

06.September 2019 -- 11:30

Title

Automatic Segmentation of MR Spinal Images Using Knowledge Transfer from CT

Student

Sebastian Lutz

Type

MA Kick-Off

Supervisor

Thomas Wendler

Additional supervisors

Magda Paschali

Director

Prof. Nassir Navab

Abstract

This thesis aims to develop a method for automatic segmentation of vertebrae in MR images. Towards this goal, knowledge transfer from an existing segmentation network for CT images will be utilized. The thesis includes acquiring patient data, registaring CT and MR images and developing a deep neural network to perform the segmentation and knowledge transfer.

Date & Time

06.September 2019 -- 12:00

Title

3D Instances from a single RGB Image

Student

Peter Mortimer

Type

MA Final

Supervisor

Federico Tombari

Additional supervisors

Yida Wang

Director

Prof. Nassir Navab

Abstract

This master thesis compares different deep learning architectures for the task of 3D scene understanding from a single RGB image. The task of recovering 3D instances from a 2D image is challenging. The deep learning networks are required to estimate depth, shape, and pose of a 3D instance. This information is not directly provided in the RGB image input. We compare two different output representations for detected 3D instances.
One is a factored representation, where the instance shape is described in voxels and the instance’s pose in separate scalars for the scale, rotation, and translation. The second representation encodes the shape and pose of each detected object directly in a point cloud.
The deep learning networks in this master thesis are trained on the SUNCG dataset, a synthetic dataset containing indoor scenes of houses. The comparison of the two output representations will be done on different quantitative and qualitative aspects.

Date & Time

06.September 2019 -- 12:30

Title

Detection of Adversarial Attacks Using Feature Visualisation

Student

Viktoria Markova

Type

BA Final

Supervisor

Manuel Nickel

Additional supervisors

Federico Tombari

Director

Prof. Nassir Navab

Abstract

Convolutional Neural Networks have shown considerable performance in the last decade. They have become the quasi-standard in most computer vision. Nevertheless they remain mostly a black box when it comes to understanding how they learn. The existence of adversarial examples exacerbates the doubts that the networks, even though showing promising results, cannot be considered for serious tasks like medical image analysis or autonomous driving.
In recent years feature visualisation techniques have been developed, which attempt to visualize the inner workings of networks in a human-interpretable way and thus understand the mechanism of the learning. This work explores whether these techniques can be utilized for the detection of adversarial attacks.

Date & Time

06.September 2019 -- 13:00

Title

A Deep Learning Approach for Monitoring the Critical View of Safety in cholecystectomy Procedures

Student

Taha Emre

Type

MA Final

Supervisor

Prof. Dr. Nassir Navab

Additional supervisors

Prof. Dr. Nicolas Padoy

Director

Prof. Dr. Nassir Navab

Abstract

Surgical societies are united in advocating for the achievement of the Critical View
of Safety (CVS) in laparoscopic cholecystectomy (LC). It is assessed by conclusive
identification of 3 biliary anatomical structures by assigning a score to each of them.
However, It is revealed that even though CVS is reported in the operative notes, in
most cases CVS could not be detected visually in the operation videos. Thus, in this
project we propose a deep learning approach for CVS assessment based on single LC
video frames, which is reliable and non-disruptive for the surgical workflow. A major
challenge we faced is lack of already available dataset. So, we created our video frame
based CVS dataset labeled by multiple expert surgeons. Additionally CVS assessment
cannot be regarded as a object detection or instance segmentation task since it is defined
on expert knowledge in laparoscopic surgery. As a result of this, It can be considered as
an image quality assessment task. In light of this, we introduce a novel self-supervised
pretraining approach based on our improved Jigsaw puzzle scheme to force the network to
focus on anatomical parts that matter in CVS. Specifically, our CNN for Jigsaw puzzle
is trained with Wasserstein loss which boosted the pretraining performance. Then we
showed the pretrained weights improved performance of our CVS scoring CNN on
single frames.

Date & Time

06.September 2019 -- 13:30

Title

Semantic Object-Level SLAM in Outdoor Environments

Student

Marcel Schmitt

Type

MA Final

Supervisor

Markus Herb

Director

PD Dr. Federico Tombari

Abstract

Reliable long term localization and mapping is a key challenge for a lot of emerging technologies, like autonomous driving. Especially in the case of visual SLAM, a technique that uses camera(s) for simultaneous localization and mapping, robust data association remains a problem over larger time intervals, since correspondences are usually established by matching local image descriptors.
This work incorporates semantic image information in order to enable long term mapping and localization, since the semantic information is arguably more invariant to changes in the environment, like season. A state of the art convolutional neural network for image segmentation is utilized to obtain semantic information of a sequence of input images. This information is used to increase robustness of a visual odometry system, that yields an initial trajectory estimation and environment reconstruction. Furthermore, the semantic information is aggregated over time, to build a semantically labeled reconstruction of the environment, i.e. a semantically labeled point cloud. The developed image-based place recognition module is able to recognize previously visited places due to their semantic appearance and proposes them as loop closure candidates. The relative pose of a loop closure candidate and the current frame is refined by a novel pure semantic reprojection error, that is modeled in a probabilistic robust manner and solver with expectation maximization. The gained information is subsequently used to perform a pose graph optimization to obtain a globally more consistent estimation of the map and the environment. Smart validation steps after each optimization ensure stable operation and reject outliers early on.
The place recognition module in combination with the subsequent pose refinement strategy show promising results and yield to relative pose constraints up to an accuracy of 0.038 meter.

Date & Time

13.September 2019 -- 11:30

Title

Dynamic Volume Estimation of Bone Cement during Vertebroplasty for Therapy Optimization

Student

Granit Bimbashi

Type

MA Kick-Off

Supervisor

Thomas Wendler

Additional supervisors

Ardit Ramadani

Director

Prof. Nassir Navab

Abstract

This master thesis is carried out with the aim of implementing a system to dynamically track the injected volume of bone cement into the vertebral body, in a procedure known as Vertebroplasty. Meeting the objective of this thesis shall satisfy the purpose of optimizing the treatment of vertebral compression fractures.

Date & Time

27.September 2019 -- 10:30

Title

Exploring CycleGANs for Anomaly Detection in Brain MRI

Student

Robert Graf

Type

BA Final

Supervisor

Christoph Baur

Director

Prof. Nassir Navab

Abstract

Recently, Autoencoders (AEs) have been successfully used for detecting anomalies in medical images. The core concept thereby is to model normal anatomy by learning to compress and reconstruct only healthy data. Anomalies in unseen data can then be detected from large reconstruction errors. For medical image analysis, this approach is quite appealing as it relieves from the need for pixel-level annotations of training images. In fact, it has been shown that this ansatz works quite well for detecting and even delineating hyper-intense lesions in brain MRI. However, working with MRI poses some challenges: MR scans acquired with devices from different manufactureres exhibit different intensity distributions, and in contrast to CT, MR images generally have no clear anatomical meaning. This variance makes it difficult for AEs to reconstruct images with high fidelity. As a workaround, we hypothesize that such anomaly detection is considerably easier on simulated data. Recent advances in unpaired image-to-image translation, i.e. CycleGANs and their siblings, might hold the potential to transform real data into its simulated counterpart. In this Bachelor's Thesis it shall be explored if CycleGANs can establish a mapping from real MR data to simulated images, on which an AE-based anomaly detection is potentially more effective. Furthermore, it shall be investigated if CycleGANs themselves can be used directly for anomaly detection by establishing a mapping from pathological data to its healthy counterpart.

Date & Time

04.Oktober 2019 -- 10:30

Title

Automatic CT Classification and Segmentation of Bone Cement after Osteoplasty for Therapy Optimization

Student

Stefan Walke

Type

MA Kick-Off

Supervisor

Thomas Wendler

Additional supervisors

Anjany Sekuboyina

Director

Prof. Nassir Navab

Abstract

Constructing a segmentation mask and estimating bone cement volume utilizing postoperative 3D CT scans, robust to non-convex shapes and accounting for partial volume effects.To integrate into the DeepSpine Pipeline a CNN based classifier to determine whether the spine has been augmented by way of bone cement or other foreign objects is build.



  • No labels