More details about this position is here:
https://www.irit.fr/~Duong-Hung.Pham/wp-content/uploads/sites/27/2024/04/Sujets_PhD_2024.pdf.

-----------------------------------------------------------------------------------------------------

• Location: Equipe MINDS, Laboratoire IRIT, Toulouse, France
• Funding : A fully-funded 3-year PhD grant has been secured (approx. 2100euros gross per month).
• Starting date: as soon as possible and no later than October 1, 2024.
• Supervisors: This Ph.D. will be supervised by:
  • Duong Hung Pham, IRIT/UPS, Toulouse, duong-hung.pham@irit.fr
  • Denis Kouame, IRIT/UPS, Toulouse, kouame@irit.fr

Keywords: ULM, transcranial imaging, super-resolution, 3D ultrasound imaging, inverse problem.

1. Description

Ultrasound Localization Microscopy (ULM) offers a solution to the conventional trade-off between spatial resolution and penetration depth by leveraging sparse microbubbles (MBs) contrast agents and ultrafast imaging [1]. The ULM process encompasses five stages: 1) data acquisition, 2) pre-processing (tissue filtering), 3) MB detection and localization, 4) MB tracking, and 5) ULM image rendering [1]. However, ULM is not without limitations:

• Aberration Correction: Pre-clinical studies often resort to craniotomies to mitigate aberrations, limiting the clinical applicability of ultrasound-based brain imaging [2].
• 2D Imaging Constraints: The prevalent use of 2D imaging restricts access to vessels within the imaging plane, potentially biasing velocity estimates and hindering the development of reliable biomarkers based on hemodynamic changes [3]. The inherent limitations of 2D ULM underscore the need for volumetric ultrasound imaging techniques.
• Localization and Acquisition Trade-off: ULM introduces a trade-off between MB localization precision and data acquisition time [4]. High MB dilution enhances localization accuracy but extends acquisition time, whereas low dilution accelerates acquisition but compromises accuracy due to signal overlap.
• Computational Efficiency: Current ULM algorithms remain time-intensive.

Efforts to tackle these challenges include the adoption of deep learning (DL) methods, such as conventional convolutional neural networks (CNNs) for point target extraction from B-mode images [5], or comprehensive CNNs for spatiotemporal US data [6]. This PhD thesis aims to address these limitations by leveraging recent advancements in inverse problems in imaging, encompassing both model-based and data-driven approaches. The goal is to enhance the performance and efficiency of transcranial ULM, especially in scenarios involving high MB concentrations. In the model-based approach, we will explore direct models that account for imaging features and various inversion scenarios. Rather than relying on black-box end-to-end deep learning methods, our emphasis is on interpretable, model-driven or physics-based deep learning techniques, such as unfolding CNNs [7]. Additionally, we will investigate non-supervised or weakly supervised approaches, including diffusion models or similar frameworks, for unsupervised deep learning training [8], [9].

2. Required Skills

• Master’s degree (or equivalent) in signal and image processing, or applied mathematics,
• Proficiency in machine learning/ deep learning techniques, with a focus on CNNs,
• Strong programming skills in either Matlab or Python,
• Proficiency in English,
• Interest in medical ultrasound imaging is beneficial but not mandatory.

3. Application

Prospective candidates should submit the following documents as a SINGLE PDF file:

• A one-page cover letter,
• Curriculum vitae, including publications if applicable,
• Two reference letters.
• The position remains open until filled. Please send all documents to duong-hung.pham@irit.fr and denis.kouame@irit.fr.

4. References

[1] B. Heiles et al., “Performance benchmarking of microbubble-localization algorithms for ultrasound localization microscopy,” en, Nat. Biomed. Eng, vol. 6, no. 5, pp. 605–616, May 2022. doi: 10.1038/s41551- 021-00824-8.

[2] P. Xing et al., “Inverse Problem Based on a Sparse Representation of Contrast-enhanced Ultrasound Data for in vivo Transcranial Imaging,” 2024. doi: 10.48550/ARXIV.2401.10389.

[3] P. Xing et al., “Towards Transcranial 3D Ultrasound Localization Microscopy of the Nonhuman Primate Brain,” 2024. doi: 10.48550/ARXIV.2404.03547.

[4] X. Chen et al., “Localization Free Super-Resolution Microbubble Velocimetry Using a Long Short-Term Memory Neural Network,” IEEE Transactions on Medical Imaging, vol. 42, no. 8, pp. 2374–2385, Aug. 2023. doi: 10.1109/TMI.2023.3251197.

[5] R. J. G. van Sloun et al., “Super-Resolution Ultrasound Localization Microscopy Through Deep Learning,” IEEE Transactions on Medical Imaging, vol. 40, no. 3, pp. 829–839, Mar. 2021. doi: 10.1109/ TMI.2020.3037790.

[6] U.-W. Lok et al., “Fast super-resolution ultrasound microvessel imaging using spatiotemporal data with deep fully convolutional neural network,” Physics in Medicine & Biology, vol. 66, no. 7, p. 075 005, Apr. 2021. doi: 10.1088/1361-6560/abeb31.

[7] V. Pustovalov et al., “Deep Unfolding RPCA for High Resolution Flow Estimation,” in 2022 EEE Int. Ultrason. Symp. (IUS), ISSN: 1948-5727, Oct. 2022, pp. 1–4. doi: 10.1109/IUS54386.2022.9957400.

[8] J. Ho et al., “Denoising Diffusion Probabilistic Models,” 2020. doi: 10.48550/ARXIV.2006.11239. [9] B. B. Moser et al., “Diffusion Models, Image Super-Resolution And Everything: A Survey,” 2024. doi: 10.48550/ARXIV.2401.00736.