Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Optoacoustic Image Reconstruction with Deep Learning
Optoacoustic tomography (OAT) is a non-ionizing imaging modality capable of acquiring high contrast and resolution images of optical absorption at depths greater than traditional optical imaging techniques. This project aims to develop algorithms to reconstruct optoacoustic images using deep learning methods.
Medical image reconstruction is an integral part of modern medicine and plays a key role in life science applications. Optoacoustic (OA) imaging as a rapidly evolving biomedical modality in specific can provide high-resolution 3D soft tissue images based on optical absorption; however, the need for rapid image formation and the practical restrictions of conventional methods that arise from the constraints of data acquisition and geometrical limitations in clinical workflow are presenting new image reconstruction challenges. To overcome these challenges and enhance the quality of reconstruction, several learning-based methods have recently been introduced for image reconstruction.
Medical image reconstruction is an integral part of modern medicine and plays a key role in life science applications. Optoacoustic (OA) imaging as a rapidly evolving biomedical modality in specific can provide high-resolution 3D soft tissue images based on optical absorption; however, the need for rapid image formation and the practical restrictions of conventional methods that arise from the constraints of data acquisition and geometrical limitations in clinical workflow are presenting new image reconstruction challenges. To overcome these challenges and enhance the quality of reconstruction, several learning-based methods have recently been introduced for image reconstruction.
The aim of the project is to analyze the learning-based methods for OA image reconstructions. Recent works cover fully learned approaches transforming the domain from signal to image, learned iterative model based, hybrid methods considering both image and signal domain data, supervised recurrent neural networks and recently unsupervised approaches. We will develop new algorithms that fits the modality-specific requirements and performs better than conventional methods in terms of time (vs. model-based reconstruction) and accuracy (vs. back projection reconstruction).
**_Required skills / what you will empower_**
- Matlab & Python
- Biomedical image analysis
- Mathematics and image processing
- ML and DL concepts
- Multidisciplinary teamwork
**Tasks (Full picture of the project):**
- Generate synthetic data similar to experimental data distribution
- Analyze and optimize preprocessing methods
- Implement and train the network
- Analyze and optimize loss functions (discarded physics could be integrated here)
- Apply domain adaptation / fine tuning from synthetic to experimental data
- Augment experimental data
- Analyze post processing methods
- Analyze uncertainty
- Evaluate robustness of the method (noise, structure, background, contrast)
- Compare results with benchmarks
The aim of the project is to analyze the learning-based methods for OA image reconstructions. Recent works cover fully learned approaches transforming the domain from signal to image, learned iterative model based, hybrid methods considering both image and signal domain data, supervised recurrent neural networks and recently unsupervised approaches. We will develop new algorithms that fits the modality-specific requirements and performs better than conventional methods in terms of time (vs. model-based reconstruction) and accuracy (vs. back projection reconstruction).
**_Required skills / what you will empower_**
- Matlab & Python
- Biomedical image analysis
- Mathematics and image processing
- ML and DL concepts
- Multidisciplinary teamwork
**Tasks (Full picture of the project):**
- Generate synthetic data similar to experimental data distribution
- Analyze and optimize preprocessing methods
- Implement and train the network
- Analyze and optimize loss functions (discarded physics could be integrated here)
- Apply domain adaptation / fine tuning from synthetic to experimental data
- Augment experimental data
- Analyze post processing methods
- Analyze uncertainty
- Evaluate robustness of the method (noise, structure, background, contrast)
- Compare results with benchmarks
Neda Davoudi (davoudin@student.ethz.ch)
PhD Candidate at ETH Zurich
Wolfgang-Pauli-Str. 27
HIT E-22/Razansky Lab
8093 Zürich
Switzerland