Subject

The principal challenge for integrating whole-body imaging in clinical routine comes from the large amount of data to be reviewed. Developments in imaging hardware continuously increase the resolution and field-of-view (FOV) of acquired images. Automatic detection and segmentation techniques could alleviate the workflow and ultimately facilitate the process of reading whole-body scans by reducing the reading time and improving the diagnostic accuracy of whole-body imaging.
In the last decade, deep learning has become the methodology of choice for many medical image processing applications. The number of publications in the field of artificial intelligence (AI) used for medical imaging is increasing exponentially year by year. Nonetheless, challenges arise when introducing deep learning in the field of medical image processing. The small data sets with noisy annotations and a large FOV create a need for non- standard deep learning strategies.
In medical imaging, due to hardware limitations, deep learning models are trained on patches rather than full images. The size of patches is often limited to 1283 voxels for 3-dimensional networks. For large images with high resolutions such as whole-body MRI, the maximum size of a patch constitutes only 1% of the total image volume. The patch-based approach is a major limitation for these applications, because the network does not have the ability to localize a patch in the image nor can it incorporate a large context for the segmentation of a patch.

Kind of work

The goal of this project is to develop methods to incorporate location information to the patches via a location embedding. Current advancements in deep learning push computer vision models from convolutional neural networks to transformer-based methods [1]. These novel transformer-based (VIT) models have an intrinsic need for patch-location embeddings. In the current state-of-the-art, simple location embeddings are used that do not incorporate any anatomical information [2]. The objective of this thesis is to develop anatomical aware location embeddings to enhance the current VIT models.

Framework of the Thesis

The developments will be performed as an extension to existing software developed within the ETRO research group to segment metastatic bone lesions from whole-body MRI. Most of the research component will focus on advanced deep learning algorithms. A strong interest in the AI and a fluency in python is heavily advised. Algorithms will be implemented in Python, using open-source deep learning libraries and image processing libraries, such as the monai [3]and SimpleITK.
Project will consist of:
• Literature study.
• Implementation of the literature-based, state-of-the-art methodology to the whole-body MRI dataset available in ETRO.
• Implementation of novel algorithm using the patch location embeddings.
• Development of a workflow allowing for quantitative and qualitative validation of obtained results.
• Optimization of the workflow by iterative improvements applied to the method and additional feedback from radiologists and engineers.
• Implementation: Python, Pytorch, Monai

References
[1] A. Hatamizadeh, V. Nath, Y. Tang, D. Yang, H. Roth, and D. Xu, “Swin UNETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images,” arXiv [eess.IV], Jan. 04, 2022
[2] A. Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” arXiv [cs.CV], Oct. 22, 2020.
[3] https://monai.io/

Number of Students

1

Expected Student Profile

The thesis focusses on advanced deep learning approaches for computer vision in the medical field. Applicants should have a strong interest towards clinical decision support and a base knowledge on (medical) image processing.

Fluency in python is heavily advised and prior experience with deep learning frameworks is a plus. Model building and testing will be done with the help of monai or pure pytorch.