Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Deep learning based 3D segmentation of X-ray microtomographic volumes from the human middle ear
The aim of the project is to develop a deep learning model capable of performing 3D semantic segmentation of the different features of the middle ear, from synchrotron-based X-ray microtomography 3D volumes. Challenges include developing a high-performance algorithm for handling large data sizes (20 GB per volume). The project consists in a first phase about dataset generation followed by a second phase of development/selection of the most appropriate DL model and metrics.
Keywords: deep learning, 3D semantic segmentation, X-ray microtomography, human middle ear, ossicular chain, synchrotron, biomedical imaging
This project is part of a larger research focus investigating the biomechanics of the human middle ear, using dynamic synchrotron-based X-ray imaging.
Characterizing the vibrations of the middle ear ossicles (malleus, incus, stapes) during sound transmission is a focal point in clinical research. However, the small size of the structures, their micrometer-scale movement, and the deep-seated position of the middle ear within the temporal bone make these types of measurements extremely challenging.
Here, we use retrospectively-gated time-resolved microtomography to obtain a 3D volume of the full middle ear system for each phase of the periodic motion. The volumes are subsequently correlated and registered to analyze the motion patterns and quantify the transformations and displacements of regions of interest.
The aim of the project proposed here is to develop / select the most appropriate deep learning model capable of performing a 3D semantic segmentation of the different features of the middle ear, from any microtomography 3D volumes. The ultimate goal is to automatically identify and segment each object of interest prior to the volume correlation and subsequent calculations, enabling a high data analysis throughput. One challenge will be to have a performant algorithm on large-size data (20 GB per volume). The subsequent segmented data obtained with the deep learning model will also benefit visualization and static anatomical investigations across biological disparity, and may serve as input data for the refinement of finite-element models of the human auditory system.
The project will be split in two phases:
1- Dataset generation & 2- DL model for 3D semantic segmentation
Phase 1 consists in assembling a dataset large enough for training and testing. A multi-ROIs labeling will have to be performed on the microtomography 3D volumes from 12 independent samples. 3D slicer, Dragonfly or Ilastik may be considered for this purpose. In case the dataset is not large enough for supervised learning, data augmentation techniques will be used, as well as potential generation of synthetic data. A semi-supervised learning may also be considered.
Phase 2 consists in the development/application of a deep learning model for the 3D semantic segmentation of the human middle ear. A baseline model will be chosen based on current literature review. Different methods can be considered (a CNN approach like 3D U-Net may be a good approach). The model will be trained and tested on the datasets from phase 1, and appropriate metrics will be used to evaluate the performance of the model.
References
[1] He, Yong, et al. "Deep learning based 3D segmentation: A survey." arXiv preprint arXiv:2103.05423 (2021).
[2] Mahmud, Bahar Uddin, et al. "Deep learning-based segmentation of 3D volumetric image and microstructural analysis." Sensors 23.5 (2023): 2640.
[3] Zhang, Shaokun, et al. "The Segmentation of Knee MR Image Using Deep Networks and Pruning Strategy." 2023 16th International CISP-BMEI. IEEE, 2023.
[4] Kayalibay, Baris, Grady Jensen, and Patrick van der Smagt. "CNN-based segmentation of medical imaging data." arXiv preprint arXiv:1701.03056 (2017).
[5] Çiçek, Özgün, et al. "3D U-Net: learning dense volumetric segmentation from sparse annotation." MICCAI 2016: 19th International Conference, Proceedings, Part II 19. Springer International Publishing, 2016.
This project is part of a larger research focus investigating the biomechanics of the human middle ear, using dynamic synchrotron-based X-ray imaging. Characterizing the vibrations of the middle ear ossicles (malleus, incus, stapes) during sound transmission is a focal point in clinical research. However, the small size of the structures, their micrometer-scale movement, and the deep-seated position of the middle ear within the temporal bone make these types of measurements extremely challenging. Here, we use retrospectively-gated time-resolved microtomography to obtain a 3D volume of the full middle ear system for each phase of the periodic motion. The volumes are subsequently correlated and registered to analyze the motion patterns and quantify the transformations and displacements of regions of interest. The aim of the project proposed here is to develop / select the most appropriate deep learning model capable of performing a 3D semantic segmentation of the different features of the middle ear, from any microtomography 3D volumes. The ultimate goal is to automatically identify and segment each object of interest prior to the volume correlation and subsequent calculations, enabling a high data analysis throughput. One challenge will be to have a performant algorithm on large-size data (20 GB per volume). The subsequent segmented data obtained with the deep learning model will also benefit visualization and static anatomical investigations across biological disparity, and may serve as input data for the refinement of finite-element models of the human auditory system.
The project will be split in two phases: 1- Dataset generation & 2- DL model for 3D semantic segmentation
Phase 1 consists in assembling a dataset large enough for training and testing. A multi-ROIs labeling will have to be performed on the microtomography 3D volumes from 12 independent samples. 3D slicer, Dragonfly or Ilastik may be considered for this purpose. In case the dataset is not large enough for supervised learning, data augmentation techniques will be used, as well as potential generation of synthetic data. A semi-supervised learning may also be considered. Phase 2 consists in the development/application of a deep learning model for the 3D semantic segmentation of the human middle ear. A baseline model will be chosen based on current literature review. Different methods can be considered (a CNN approach like 3D U-Net may be a good approach). The model will be trained and tested on the datasets from phase 1, and appropriate metrics will be used to evaluate the performance of the model.
References [1] He, Yong, et al. "Deep learning based 3D segmentation: A survey." arXiv preprint arXiv:2103.05423 (2021). [2] Mahmud, Bahar Uddin, et al. "Deep learning-based segmentation of 3D volumetric image and microstructural analysis." Sensors 23.5 (2023): 2640. [3] Zhang, Shaokun, et al. "The Segmentation of Knee MR Image Using Deep Networks and Pruning Strategy." 2023 16th International CISP-BMEI. IEEE, 2023. [4] Kayalibay, Baris, Grady Jensen, and Patrick van der Smagt. "CNN-based segmentation of medical imaging data." arXiv preprint arXiv:1701.03056 (2017). [5] Çiçek, Özgün, et al. "3D U-Net: learning dense volumetric segmentation from sparse annotation." MICCAI 2016: 19th International Conference, Proceedings, Part II 19. Springer International Publishing, 2016.
Develop a deep-learning based 3D semantic segmentation for the automatic segmentation of human middle ear features (especially tympanic membrane and the 3 bones from the ossicular chain: malleus, incus and stapes). This method will be applied on large-size synchrotron-based X-ray microtomography data.
Develop a deep-learning based 3D semantic segmentation for the automatic segmentation of human middle ear features (especially tympanic membrane and the 3 bones from the ossicular chain: malleus, incus and stapes). This method will be applied on large-size synchrotron-based X-ray microtomography data.