ETRO VUB
About ETRO  |  News  |  Events  |  Vacancies  |  Contact  
Home Research Education Industry Publications About ETRO

Master theses

Current and past ideas and concepts for Master Theses.

Semantic Segmentation of Medical Imaging via Graph Convolutional Networks

Subject

Semantic medical imaging segmentation consists of detecting and delineating regions of interest in medical images. Manual semantic segmentation implies a burden on the health care systems because it usually requires that medical staff members expend a considerable amount of time on this task, which is a tedious and fatiguing responsibility. It is also a challenging task and must be performed by medical staff members with expertise. Hence, learning systems for automatic medical imaging segmentation has caught increasing interest from academics and medical experts because it can play a fundamental role in Computer-Aided Diagnosis. In general, these learning systems are concerned with automatic semantic segmentation of medical imaging applied to different tasks, such as segmentation of lesions suggestive of some diseases or segmentation of anatomical or meaningful morphological structures [1].

Most state-of-the-art learning models build on Deep Convolutional Neural Networks (DCNNs) that have proved outstanding performance in several medical imaging semantic segmentation tasks [1]. In DCNNs' architectures, the Convolutional layers are combined with Pooling layers. However, these layers' receptive fields cannot capture the full range of spatial and across channel interdependencies, and pooling operations entail a considerable loss of contextual and location information, which is not desirable in semantic segmentation tasks. Researchers have proposed alternatives to remedy such issues, such as using Dilated Convolutions or Self-Attention modules in conjunction with multiscale features fusion that can enhance the DCNNS towards exploiting more properly contextual and location information [2, 3]. Nonetheless, not enough attention has been devoted to semantic segmentation of medical imaging using Graph Convolutional Neural Networks (GCNNs) to enhance models' representation power considering long-range contextual and location information interdependencies [4, 5].

GCNNs are emerging models with growing popularity in the Machine Learning community because of their expressive capacity to model relationships and interdependencies on multi-dimensional data [6]. Previous studies suggest that GCNNs hold properties that can model spatial and across channel relationships and interdependencies [4, 5]. They have also proven effective in semantic segmentation and scene understanding tasks [4, 5]. Hence, they can equip the semantic segmentation models for exploiting contextual and location information to attain greater representational power [7].

1. Isensee, Fabian, et al. "nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation." Nature Methods 18.2 (2021): 203-211.
2. Chen, Liang-Chieh, et al. "Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs." IEEE transactions on pattern analysis and machine intelligence 40.4 (2017): 834-848.
3. Fu, Jun, et al. "Dual attention network for scene segmentation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019
4. Zhang, Li, et al. "Dual graph convolutional network for semantic segmentation." arXiv preprint arXiv:1909.06121 (2019).
5. Liu, Qinghui, et al. "SCG-Net: Self-Constructing Graph Neural Networks for Semantic Segmentation." arXiv preprint arXiv:2009.01599 (2020).
6. Zhang, Si, et al. "Graph convolutional networks: a comprehensive review." Computational Social Networks 6.1 (2019): 1-23.
7. Lu, Yi, et al. "CNN-G: convolutional neural network combined with graph for image segmentation with theoretical analysis." IEEE Transactions on Cognitive and Developmental Systems (2020).

Kind of work

This thesis will deal with semantic segmentation of Computed Tomography images. The research project will investigate recent trends in Deep Learning techniques to develop an automatic semantic segmentation model for medical imaging based on GCNNs, focusing on exploiting contextual and location information by modeling features interdependencies. We will validate the developed model on different sematic segmentation tasks such as lung or retinal vessel segmentation and semantic segmentation of lesions related to defined diseases. The student will (i) investigate different state-of-the-art approaches for semantic image segmentation based on GCNNs and identify the gaps in the literature, (ii) formulate the graph-based problem for semantic medical imaging segmentation, (iii) develop a model for sematic medical imaging segmentation based on GCNNs, and (iv) validate the developed model on the publicly/privately available data.

Framework of the Thesis

AI-based chest CT analysis enabling rapid COVID diagnosis and prognosis (icovid)

Number of Students

1

Expected Student Profile

• Experience with image and signal processing.
• Experience with machine learning and statistics.
• Strong programming skills (Python).
• Interest in developing state-of-the-art Machine Learning methods and conduct experiments.
• Ability to write scientific reports and communicate research results at conferences in English.

Promotor

Prof. Hichem Sahli

+32 (0)2 629 2916

hsahli@etrovub.be

more info

Supervisor

Mr. Abel Díaz Berenguer

+32 (0)2 629 1029

aberengu@etrovub.be

more info

- Contact person

- IRIS

- AVSP

- LAMI

- Contact person

- Thesis proposals

- ETRO Courses

- Contact person

- Spin-offs

- Know How

- Journals

- Conferences

- Books

- Vacancies

- News

- Events

- Press

Contact

ETRO Department

Tel: +32 2 629 29 30

©2022 • Vrije Universiteit Brussel • ETRO Dept. • Pleinlaan 2 • 1050 Brussels • Tel: +32 2 629 2930 (secretariat) • Fax: +32 2 629 2883 • WebmasterDisclaimer