Upcoming oberseminars

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

Location: MI 03.13.010 (Seminar Room)

Zoom Link: The link is shared with CAMP members via email roughly two days before each presentation. (To students: please ask your project supervisors for the Zoom link)

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.

	Presentation	Discussion
Thesis Kick-off	20 min	10 min
Thesis Final	25 min	5 min
IDP Kick-off	10 min	10 min
IDP Final	15 min	5 min
Guided Research Final	15 min	5 min

Schedule:

Date	Time	Presenter	Title	Type	Supervisor/Contact
31.Mär (Fr)	10:30	Markus Hamberger	4D OCT and Enface fusion for enhanced surgical visualization	MA Kick-Off	Michael Sommersperger
31.Mär (Fr)	11:00	Miguel Hoyos	Exploring the Magic Mirror for Anatomy Learning on Wearable Head-Mounted Displays	MA Kick-Off	Dr. Ulrich Eck
31.Mär (Fr)	11:30	Mahaut Gerard	Augmentation techniques for few-shot learning	MA Final	Azade Farshad
31.Mär (Fr)	12:00	Oleksandra Adonkina	Pixel-Level Explanation of Multiple Instance Learning Models in Biomedical Single Cell Images	MA Final	Ario Sadafi
31.Mär (Fr)	12:30	Iuliia Skobleva	Driver Pose Estimation and Identification of Seating Position Anomalies	MA Final	Patrick Ruhkamp
05.Mai (Fr)	10:30	Leonhard Zirus	Segmentation of Nerves for patients with Lepra	BA Kick-Off	Vanessa Gonzalez
05.Mai (Fr)	11:00	Simay Sanli	Analysis of Performance Metrics for Surgical Phase Recognition: An online tool for insights and analysis	MA Kick-Off	Felix Holm
05.Mai (Fr)	11:30	Hongcheng Zhao	Category-Level Pose Estimation in Cluttered Scenes	MA Kick-Off	Pengyuan Wang
05.Mai (Fr)	12:00	Pavel Jahoda	Progressive Semantic Image Manipulation Using Objectwise Features	MA Final	Azade Farshad
26.Mai (Fr)	10:30	Shruti Pistolwala	Fusing Spatially Consistent Past and Present Real-time 3D Reconstruction for Immersive Medical Teleconsultation	MA Final	Kevin Yu

Detailed schedule:

Date & Time	31.März 2023 -- 10:30
Title	4D OCT and Enface fusion for enhanced surgical visualization
Student	Markus Hamberger
Type	MA Kick-Off
Supervisor	Michael Sommersperger
Director	Prof. Nassir Navab
Abstract	Swept-source Optical Coherence Tomograph (OCT) has enabled near video-rate threedimensional imaging for ophthalmic procedures. While 4D OCT provides many advantages in depth-resolved visualization of surgical maneuvers, features from conventional imaging, such as the microscopic view are not present in 4D OCT. To maximize the information provided to the surgeon, image fusion of these modalities is a promising research direction to achieve rich surgical visualizations and perception during challenging microsurgical tasks.

Date & Time	31.März 2023 -- 11:00
Title	Exploring the Magic Mirror for Anatomy Learning on Wearable Head-Mounted Displays
Student	Miguel Hoyos
Type	MA Kick-Off
Supervisor	Dr. Ulrich Eck
Director	Nassir Navab
Abstract	In recent decades, there has been a reduction in gross anatomy teaching time within the medical curriculum in universities. This has created different interactive, immersive systems for active learning in anatomy education. One such system is the Magic Mirror, which allows the visualization of different human body systems through an AR Mirror over the user’s own body using a large screen display and a Microsoft Kinect RGB-D Sensor. Despite having demonstrated advantages in the learning process, such as increased student motivation, improved knowledge retention, and a high level of engagement, the Magic Mirror has some drawbacks. Examples are the difficulty of controlling the AR Window in the upper limbs, visualizing the anatomy of the back, and the reduced space in which the user can interact. This study aims to overcome these disadvantages by exploring the Magic Mirror on wearable head-mounted displays like the Hololens in combination with room-scale body-tracking using multiple Microsoft Azure Kinect cameras. This occurs in a tracking room where one or multiple students wear the Hololens, and the Azure Kinects track the other students. The students wearing the Hololens, looking at the others in the room, can control the AR Window with their gaze. The AR Window can be locked and moved from one tracked student to another to visualize different virtual human body systems. In addition, they can interact with a different virtual button using his hand. To complement this system and have a better learning process, a new mode is added, the Learning Mode. In this case, the students wearing the Hololens can visualize a full-size virtual human model and its respective body systems. Additional information is displayed when interacting with each human body system.

Date & Time	31.März 2023 -- 11:30
Title	Augmentation techniques for few-shot learning
Student	Mahaut Gerard
Type	MA Final
Supervisor	Azade Farshad
Additional supervisors	Yousef Yeganeh, Etienne Bennequin,
Director	Prof. Navab
Abstract	Few-shot learning methods are built to work with few labeled samples. The performance of these methods highly depends on the quality and representativeness of the few labeled samples, all the more since the number of seen samples is small. In Few-Shot Image Classification, we classify incoming images among a set of classes, with no prior information about these classes except a few labeled images that constitute the support set. We use several support set data augmentation techniques: geometric transforms, color transforms, and a style transfer transform, to leverage the choice importance of each support set image. We compare their performance impact for three different few-shot classifiers, several sets of hyperparameters changing for example the distance, and two datasets. Besides, we observe that the euclidean distance is more sensitive to support set augmentation than the cosine distance. We show that some support set augmentation strategies can indeed greatly increase classification performance. Our work provides insights into which support set augmentation techniques are efficient to counter the lack of labeled data in few-shot image classification.

Date & Time	31.März 2023 -- 12:00
Title	Pixel-Level Explanation of Multiple Instance Learning Models in Biomedical Single Cell Images
Student	Oleksandra Adonkina
Type	MA Final
Supervisor	Ario Sadafi
Director	prof. Nassir Navab
Abstract	Explainability is a key requirement for computer-aided diagnosis systems in clinical decision-making. Multiple instance learning with attention pooling provides instance-level explainability, however for many clinical applications a deeper, pixel-level explanation is desirable, but missing so far. In this work, we investigate the use of four attribution methods to explain a multiple instance learning models: GradCAM, Layer-Wise Relevance Propagation (LRP), Information Bottleneck Attribution (IBA), and InputIBA. With this collection of methods, we can derive pixel-level explanations on for the task of diagnosing blood cancer from patients’ blood smears. We study two datasets of acute myeloid leukemia with over 100 000 single cell images and observe how each attribution method performs on the multiple instance learning architecture focusing on different properties of the white blood single cells. Additionally, we compare attribution maps with the annotations of a medical expert to see how the model's decision-making differs from the human standard. Our study addresses the challenge of implementing pixel-level explainability in multiple instance learning models and provides insights for clinicians to better understand and trust decisions from computer-aided diagnosis systems.

Date & Time	31.März 2023 -- 12:30
Title	Driver Pose Estimation and Identification of Seating Position Anomalies
Student	Iuliia Skobleva
Type	MA Final
Supervisor	Patrick Ruhkamp
Additional supervisors	Benjamin Busam + MIT AgeLab Partners
Director	Nassir Navab
Abstract	Ensuring driver’s safety is the most important component of developing novel vehicle technologies. As the car is becoming capable to operate with little to no human intervention, people will start diverging their attention from driving and focus on more enjoyable tasks such as reading, being on the phone or playing with the entertainment systems. It is during these breaks that people will very likely be out-of-position, which may pose increased risks on them as the current safety features, installed in most cars, were developed for people sitting straight with their hands on the wheel. We develop a method to understand the frequency and types of out-of-position postures and how they correlate with anatomy levels. We propose to estimate the pose of the driver using a 2D top-down pose estimation algorithm and use the results of those predictions to estimate anomalies using Cluster-Based Local Outlier Factor. We identify common types of anomalies as well as find that out-of-position sequences on average are more likely to happen in L2, which corresponds to the "hands-off" level of autonomy, than in manual mode.

Date & Time	05.Mai 2023 -- 10:30
Title	Segmentation of Nerves for patients with Lepra
Student	Leonhard Zirus
Type	BA Kick-Off
Supervisor	Vanessa Gonzalez
Director	Prof. Nassir Navab
Abstract	Lepra participants suffer from nerve degeneration due to the disease. Nowadays, It does not exist follow-up method for evaluating the loss of nerve fibers. Our objective is to provide an automatic Nerve tracking method. It will segment automatically the Cross-sectional area of 6 bilateral views of the nerves in US videos. It will be trained on 100 participants, extracting 10 for testing.

Date & Time	05.Mai 2023 -- 11:00
Title	Analysis of Performance Metrics for Surgical Phase Recognition: An online tool for insights and analysis
Student	Simay Sanli
Type	MA Kick-Off
Supervisor	Felix Holm
Additional supervisors	Tobias Czempiel
Director	Prof. Nassir Navab
Abstract	Surgical Phase Recognition is a significant and challenging task that has the potential to enhance patient safety and become an important component of computer-assisted surgical systems. However, due to various challenges in real-world surgical videos such as longer durations, the occurrence of only some phases during surgeries and multiple video sources, metric selection and accurate evaluation can be difficult. The choice of suitable metrics for this task depends on the specific applications and the relative importance of the different error types. For instance, in some surgical procedures with phases of varying importance, the macro-average scoring metric may be more relevant than the micro-average scoring metric. Additionally, the presence of imbalanced classes can introduce an inconsistency in the calculation of micro-average and macro-average metrics, as it can lead to the dominance of certain classes over others. In this thesis, our goal is to develop a web application that automates metric calculations, as well as label and prediction edits, for both single and multiple surgical videos. The web application will also provide explanations of metrics and facilitate the verification of metric results. We recommend potentially suitable metrics and identify possible pitfalls of current metrics for various scenarios and applications.

Date & Time	05.Mai 2023 -- 11:30
Title	Category-Level Pose Estimation in Cluttered Scenes
Student	Hongcheng Zhao
Type	MA Kick-Off
Supervisor	Pengyuan Wang
Director	Dr. Benjamin Busam
Abstract	In recent years, category-level object pose estimation has attracted many researchers’ attention due to its ability to predict the 6D pose and 3D metric size for previously unseen objects. Existing methods mostly predicts the pose for the RGB-D input observation of each individual object. These methods work relatively well for sparsely organized objects, but particularly struggle with cluttered scenes where object interactions such as occlusions and contacts are present, rendering the input observations less complete and the pose prediction task more challenging. Therefore, in this thesis, we aim to exploit the layout information in the RGB-D image which intuitively could improve the performance of category-level pose prediction in cluttered scenes. To this end, we first reconstruct target object using an implicit representation. Second, we introduce the relationship loss that regularizes the relationships between objects given their shape features and pose predictions. Third, we also introduce the occlusion loss to avoid the violation of scene context. Our approach could have applications in real-world scenarios such as robotics and augmented reality.

Date & Time	05.Mai 2023 -- 12:00
Title	Progressive Semantic Image Manipulation Using Objectwise Features
Student	Pavel Jahoda
Type	MA Final
Supervisor	Azade Farshad
Director	Prof. Nassir Navab
Abstract	Semantic image manipulation can be performed by conditioning an image generation model on a source image and the target segmentation map or semantic scene graph. Manipulating scenes using scene graphs make it easier for the user to add or remove objects or change the relationships between the objects by simply changing the nodes and edges in the graph. However, the scene graph-to-image model is limited to generating low-resolution images. The possibility of cropping parts of the image and conditioning the model on cropped parts of the segmentation map makes it possible to generate high-resolution manipulated images. Nevertheless, conditioning the image generator model on a small part of an image would make a scene graph sparse and would not convey enough information for generating the object with high-resolution background information. In this thesis, we explore ways to condition an image generation model on parts of a scene graph without facing the challenge of sparse graphs.

Date & Time	26.Mai 2023 -- 10:30
Title	Fusing Spatially Consistent Past and Present Real-time 3D Reconstruction for Immersive Medical Teleconsultation
Student	Shruti Pistolwala
Type	MA Final
Supervisor	Kevin Yu
Additional supervisors	Ulrich Eck
Director	Nassir Navab
Abstract	ArtekMed is an asymmetric and immersive teleconsultation VR/AR system for emergency scenarios. For example, even inexperienced paramedics may be deployed in a mass-casualty accident. Similarly, in emergency ORs, doctors may encounter rare cases requiring a remote expert's consultation. In both cases, the remote expert will likely join late to the treatment. ArtekMed captures the live-point cloud of the emergency site and stores the data in recordings. However, how can we visualize the 3D recording of the past to the remote expert and allow interaction with temporal data intuitively without disconnecting them from the live scene? In order to tackle this missing gap, we present our method of Fusing Past and Present in 3D Telepresence to merge recordings with the live 3D reconstruction of the environment. In particular, we want to deploy a virtual wire-frame cube, where the space within it is replaced by a past recording of the same location with its original scale rather than the live data. The thesis will use the approach of Magnoramas/World in Miniatures; however, it focuses on handling temporal data in addition to spatial data. We further also want to compare different visualization techniques by implementing different shaders. Finally, for a user study, we investigate how quickly users solve a task with different visualization techniques.

Space shortcuts

Page tree

Detailed schedule: