Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Automated fetal lung segmentation
Fetal magnetic resonance imaging (MRI) has become a widely available and clinically valuable MRI technique allowing earlier diagnosis of congenital defects and improved prenatal planning. Iit is critically important to be able to reliably measure the fetal lung volume, particularly in cases of pathologies impacting the lungs. Our research project (either as a semester or master thesis project) aims to develop a novel, deep-learning based image segmentation method that is able to segment the fetal lung based on super-resolution MRI data.
Keywords: magnetic resonance imaging; deep learning; machine learning; fetal development
Fetal magnetic resonance imaging (MRI) has become a widely available and clinically valuable MRI technique allowing earlier diagnosis of congenital defects and improved prenatal planning. It is critically important to be able to reliably measure the fetal lung volume, particularly in cases of pathologies impacting the lungs (1). The manual delineation of both lungs in a stack of fetal MRI images is time consuming and prone to observer error. To overcome this limitation, automated methods free of observer bias are necessary.
Our team has led important international initiatives that used ma-chine learning for the automated segmentation of fetal brain structures (2). As a next step, we aim to translate similar techniques to other vitally important anatomical structures, such as the lungs. The research project relies on fetal MRI data of over 100 fetuses, reconstructed using a super-resolution technique into an isotropically scaled 3D image (3). Three radiologists have manually outlined the lungs, which annotation can serve as ground truth data for training a machine learning algorithm (e.g. 2D or 3D U-Net). The evaluation of the performance will be possible by comparing the overlap and distance to the ground truth, as well as to evaluate the ability to reliably estimate lung volumes.
Requirements:
- A bachelor’s degree in physics, engineering, computer science, or related
- Documented experience in deep learning techniques and using machine learning frameworks (e.g. TensorFlow, Scikit-learn, Pytorch, Keras)
- Ideally some background in medical image analysis
- Good knowledge of English
1. Szpinda M, Siedlaczek W, Szpinda A, Woźniak A, Mila-Kierzenkowska C, Wiśniewski M. Volumetric growth of the lungs in human fetuses: an anatomical, hydrostatic and statistical study. Surg Radiol Anat 2014;36:813–820 doi: 10.1007/s00276-014-1269-7.
2. Payette K, de Dumast P, Kebiri H, Ezhov I, Paetzold JC, Shit S, Iqbal A, Khan R, Kottke R, Grehten P, Ji H, Lanczi L, Nagy M, Beresova M, Nguyen TD, Natalucci G, Karayannis T, Menze B, Bach Cuadra M, Jakab A. An automatic multi-tissue human fetal brain segmentation benchmark using the Fetal Tissue Annotation Dataset. Sci Data. 2021 Jul 6;8(1):167. doi: 10.1038/s41597-021-00946-3.
3. Uus A, Zhang T, Jackson LH, et al. Deformable Slice-to-Volume Registration for Motion Correction of Fetal Body and Placenta MRI. IEEE Trans Med Imaging 2020;39:2750–2759 doi: 10.1109/TMI.2020.2974844.
Fetal magnetic resonance imaging (MRI) has become a widely available and clinically valuable MRI technique allowing earlier diagnosis of congenital defects and improved prenatal planning. It is critically important to be able to reliably measure the fetal lung volume, particularly in cases of pathologies impacting the lungs (1). The manual delineation of both lungs in a stack of fetal MRI images is time consuming and prone to observer error. To overcome this limitation, automated methods free of observer bias are necessary.
Our team has led important international initiatives that used ma-chine learning for the automated segmentation of fetal brain structures (2). As a next step, we aim to translate similar techniques to other vitally important anatomical structures, such as the lungs. The research project relies on fetal MRI data of over 100 fetuses, reconstructed using a super-resolution technique into an isotropically scaled 3D image (3). Three radiologists have manually outlined the lungs, which annotation can serve as ground truth data for training a machine learning algorithm (e.g. 2D or 3D U-Net). The evaluation of the performance will be possible by comparing the overlap and distance to the ground truth, as well as to evaluate the ability to reliably estimate lung volumes.
Requirements: - A bachelor’s degree in physics, engineering, computer science, or related - Documented experience in deep learning techniques and using machine learning frameworks (e.g. TensorFlow, Scikit-learn, Pytorch, Keras) - Ideally some background in medical image analysis - Good knowledge of English
1. Szpinda M, Siedlaczek W, Szpinda A, Woźniak A, Mila-Kierzenkowska C, Wiśniewski M. Volumetric growth of the lungs in human fetuses: an anatomical, hydrostatic and statistical study. Surg Radiol Anat 2014;36:813–820 doi: 10.1007/s00276-014-1269-7.
2. Payette K, de Dumast P, Kebiri H, Ezhov I, Paetzold JC, Shit S, Iqbal A, Khan R, Kottke R, Grehten P, Ji H, Lanczi L, Nagy M, Beresova M, Nguyen TD, Natalucci G, Karayannis T, Menze B, Bach Cuadra M, Jakab A. An automatic multi-tissue human fetal brain segmentation benchmark using the Fetal Tissue Annotation Dataset. Sci Data. 2021 Jul 6;8(1):167. doi: 10.1038/s41597-021-00946-3.
3. Uus A, Zhang T, Jackson LH, et al. Deformable Slice-to-Volume Registration for Motion Correction of Fetal Body and Placenta MRI. IEEE Trans Med Imaging 2020;39:2750–2759 doi: 10.1109/TMI.2020.2974844.
1. Train and evaluate a deep neuronal network for fetal lung segmentation
2. Incorporate uncertainty of predictions into the network based on ground truth information from three annotators
3. Prepare a report, contribute to a conference publication or journal article based on the results
1. Train and evaluate a deep neuronal network for fetal lung segmentation 2. Incorporate uncertainty of predictions into the network based on ground truth information from three annotators 3. Prepare a report, contribute to a conference publication or journal article based on the results
PD Dr. med. Andras Jakab, PhD
andras.jakab@kispi.uzh.ch
Center for MR-Research
University Children’s Hospital Zürich
PD Dr. med. Andras Jakab, PhD andras.jakab@kispi.uzh.ch Center for MR-Research University Children’s Hospital Zürich