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.

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.



PresentationDiscussion
Thesis Kick-off20 min10 min
Thesis Final25 min

5 min

IDP Kick-off10 min10 min
IDP Final15 min5 min
Guided Research Final15 min5 min


Schedule:



Date

Time

Presenter

Title

Type

Supervisor/Contact

14.Okt (Fr) 10:30 Patris Valera / Zitti Andrei Multi-Label Classification in Operating Rooms klinisches Anwendungsfach Ege Özsoy
14.Okt (Fr) 11:00 Anna Cocan Multi-Label Classification in Operating Rooms klinisches Anwendungsfach Ege Özsoy
14.Okt (Fr) 11:30 Robin Frasch Enhancing multi-organ segmentation using label-dependency BA Final Dr Thomas Wendler
14.Okt (Fr) 12:30 Oleksandra Adonkina Pixel-wise explanation of multiple instance learning neural networks for biomedical single cell images MA Kick-Off carsten.marr@helmholtz-muenchen.de
04.Nov (Fr) 10:30 Ana-Maria Lacatusu Operating Room Re-identification IDP Final Tobias Czempiel
04.Nov (Fr) 11:00 Festina Ismali GAN-Based Data Augmentation for Imbalance Problems in Body Part X-Ray Classification MA Final Matthias Keicher
04.Nov (Fr) 11:30 Siyue Meng End to End Image Synthesis and Catheter Localization Using Generative Adversarial Networks MA Kick-Off Shahrooz Faghihroohi
04.Nov (Fr) 12:00 Miruna Gafencu Vertebrae shape completion in ultrasound MA Kick-Off Mohammad Farid Azampour
04.Nov (Fr) 12:30 Yevhenni Patrushev Spatial position prediction of the ultrasound probe using ORB-SLAM Guided Research Vanessa GONZALEZ
11.Nov (Fr) 10:30 Pauline Kister Implementation of a MONAI-Label Application for IFSS-Net Guided Research Vanessa Gonzalez
11.Nov (Fr) 11:00 Martin Axelrad Deep Reinforcement Learning for Facet Joint Detection in Robotic Ultrasound MA Final Maria Tirindelli
11.Nov (Fr) 11:30 Leonhard Zirus Monailabel app for DeepAtlas klinisches Anwendungsfach Vanessa GONZALEZ


Detailed schedule:



Date & Time

14.Oktober 2022 -- 10:30

Title

Multi-Label Classification in Operating Rooms

Student

Patris Valera / Zitti Andrei

Type

klinisches Anwendungsfach

Supervisor

Ege Özsoy

Director

Prof. Nassir Navab

Abstract

Surgical Operations, which take place in operating rooms of different sizes, require the use of various medical instruments. These medical instruments are numerous and are moved a lot during the operations performed by the medical staff. The 4D-OR dataset includes 10 mock surgeries, where 6 cameras were used to capture the operating room. In this project, the goal is to use the captured images to analyse the presence of instruments in the scene using computer vision.

Date & Time

14.Oktober 2022 -- 11:00

Title

Multi-Label Classification in Operating Rooms

Student

Anna Cocan

Type

klinisches Anwendungsfach

Supervisor

Ege Özsoy

Director

Prof. Nassir Navab

Abstract

Surgical Operations, which take place in operating rooms of different sizes, require the use of various medical instruments. These medical instruments are numerous and are moved a lot during the operations performed by the medical staff. The 4D-OR dataset includes 10 mock surgeries, where 6 cameras were used to capture the operating room. In this project, the goal is to use the captured images to analyse the presence of instruments in the scene using computer vision.

Date & Time

14.Oktober 2022 -- 11:30

Title

Enhancing multi-organ segmentation using label-dependency

Student

Robin Frasch

Type

BA Final

Supervisor

Dr Thomas Wendler

Additional supervisors

Francesca De Benetti

Director

Prof Nassir Navab

Abstract

TBD

Date & Time

14.Oktober 2022 -- 12:30

Title

Pixel-wise explanation of multiple instance learning neural networks for biomedical single cell images

Student

Oleksandra Adonkina

Type

MA Kick-Off

Supervisor

carsten.marr@helmholtz-muenchen.de

Additional supervisors

Ario Sadafi, Ashkan Khakzar

Director

Prof. Nassir Navab

Abstract

TBA

Date & Time

04.November 2022 -- 10:30

Title

Operating Room Re-identification

Student

Ana-Maria Lacatusu

Type

IDP Final

Supervisor

Tobias Czempiel

Additional supervisors

Lennart Bastian

Director

Nassir Navab

Abstract

TBA

Date & Time

04.November 2022 -- 11:00

Title

GAN-Based Data Augmentation for Imbalance Problems in Body Part X-Ray Classification

Student

Festina Ismali

Type

MA Final

Supervisor

Matthias Keicher

Additional supervisors

Shahrooz Faghihroohi

Director

Prof. Dr. Nassir Navab

Abstract

X-ray imaging in Digital Imaging and Communications in Medicine (Digital Imaging
and Communications in Medicine ( DICOM )) [12] format is the standard imaging
modality in healthcare information systems. The DICOM standard consists of headers
that describe the image, including the body part under examination. However, this
information can either be optional, unstructured or ambiguous, leading up to a 10%
error rate in categorizing image data [5]. An automatic system for accurate classification
of body parts from X-ray scans can fill this lack. However, publicly available x-ray
datasets with different body parts suffer from severe class imbalance, where the number
of samples in one body part is significantly lower than in others. Furthermore, some
body parts can be quite similar, making it easier for one class to be mistaken for the other.
The goal of this master thesis is to research and evaluate the current state of the art in
generative adversarial networks Generative Adversarial Networks (GAN)s as a method
of synthetic data augmentation to achieve a higher downstream performance. Moreover,
the class boundary problem is explored and GAN s that can generate hard samples are
investigated. Finally, it is demonstrated that applying GAN bases augmentations to
our implemented baseline classifier can be beneficial compared to just using standard
augmentations. However, it will be shown that these types of datasets are rather
challenging to work with, especially is separating the decision boundaries between
different classes.

Date & Time

04.November 2022 -- 11:30

Title

End to End Image Synthesis and Catheter Localization Using Generative Adversarial Networks

Student

Siyue Meng

Type

MA Kick-Off

Supervisor

Shahrooz Faghihroohi

Director

Prof. Nassir Navab

Abstract

Using low-dosage X-ray images for catheter localization would reduce the potential risk of radiation. However, it would be more challenging to segment the catheter tip due to the low contrast and SNR. Training a deep learning model would also need a huge amount of labeled low dosage data.
Making low-dose synthetic images from high-dose data can be one possible remedy for this problem.
In this project, we attempt to frame an end-to-end X-ray image synthesis and catheter segmentation model. We use generative adversarial networks to synthesize low dosage images from high dosage images, feed them into segmentation networks and
make use of the segmentation results to generate better synthetic low dosage images, in the end achieving jointly image synthesis and segmentation.

Date & Time

04.November 2022 -- 12:00

Title

Vertebrae shape completion in ultrasound

Student

Miruna Gafencu

Type

MA Kick-Off

Supervisor

Mohammad Farid Azampour

Additional supervisors

Dani Velikova, Mahdi Saleh, Thomas Wendler

Director

Prof. Dr. Nassir Navab

Abstract

Facet joint insertion is a medical procedure in which a numbing agent and/or cortisone is injected in the small joints in between the vertebrae to alleviate and locate back pain. Currently it is performed under X-Ray or Computer Tomography (CT), which leads to radiation both for the patient and the medical staff. Moreover, these modalities can only be used to verify the correct location of the needle before the injection is performed. In contrast, ultrasound (US) is a cost-effective, mobile and radiation-free alternative modality which, additionally would enable live guidance during infiltration. The introduction of US for facet joint insertion is conditioned by a good visualisation of the vertebrae which is inherent to the above-mentioned X-Ray based methods. Due to the high impedance of the bone, a spine US displays shadows below the bone surface rendering the lower part of the vertebrae invisible. A first step towards displaying the complete vertebrae in the US scan, e.g through augmented reality (AR) techniques is the ability to complete the partially visible shape of each vertebrae. Therefore, the purpose of this project is to train a deep learning (DL) model to internally represent the shape of the diverse vertebrae obtained from CT and realistically complete partial shapes in the US scans. Another application of vertebrae shape completion is for the optimal registration of the preoperative CT images with the intraoperative US images which could enable live tool tracking in the CT image.

Date & Time

04.November 2022 -- 12:30

Title

Spatial position prediction of the ultrasound probe using ORB-SLAM

Student

Yevhenni Patrushev

Type

Guided Research

Supervisor

Vanessa GONZALEZ

Director

Profesor Nassir Navab

Abstract

Volume reconstruction quality on 3d freehand ultrasound highly depends on the correct prediction of the position of the probe. Methods to predict probe position purely based on B-mode images struggle with angular movements.
As a solution, we propose to use ORB-SLAM method on natural images of the camera attached to the probe to increase the prediction of the angle.
The objective of this project is to create the camera dataset and test positioning image methods.

Date & Time

11.November 2022 -- 10:30

Title

Implementation of a MONAI-Label Application for IFSS-Net

Student

Pauline Kister

Type

Guided Research

Supervisor

Vanessa Gonzalez

Director

Nassir Navab

Abstract

The goal of this project is to create and evaluate a MONAI-Label application that allows for easy segmentation of ultrasound volumes with IFSS-Net.
During the process:
1. The existing implementation of IFSS-Net will be submitted as a contribution to the MONAI model zoo
2. A MONAI-Label application will be implemented that allows for the usage of IFSS-Net in Imfusion.
3. The decremental learning process of IFSS-Net will be evaluated with the help of medical experts

Date & Time

11.November 2022 -- 11:00

Title

Deep Reinforcement Learning for Facet Joint Detection in Robotic Ultrasound

Student

Martin Axelrad

Type

MA Final

Supervisor

Maria Tirindelli

Additional supervisors

Marco Esposito

Director

Prof. Nassir Navab

Abstract

Autonomous ultrasound is still a challenging task in which Reinforcement Learning plays a crucial role in the current development of the field. A system capable of navigate and locate complex biological structures using ultrasound images requires enough high-quality data to be trained, being this limited frequently by hardware availability. In this work, we introduce an environment that uses hybrid ultrasound simulation as a base for a simulated robotic environment, and multiple improvements over the state-of-the-art specifically designed for robotic US facet joint localization.

Date & Time

11.November 2022 -- 11:30

Title

Monailabel app for DeepAtlas

Student

Leonhard Zirus

Type

klinisches Anwendungsfach

Supervisor

Vanessa GONZALEZ

Director

Profesor Nassir Navab

Abstract

Manual segmentation of muscles in 3D ultrasound (US) sequences and volumes towards their quantitative analysis is very expensive and time-consuming.
As solution, a fully supervised multi-label segmentation methods already implemented in MONAI will be tested in our US dataset. This modality suffer from blurred contours reason why some adaptation will be made.
The objective of this project is to implement an APP for DeepAtlas method using ultrasound to ultrasound registration.






  • No labels