On January 23rd 2024 at 16.30, Ine Dirks will defend their PhD entitled “COMPUTER-AIDED DIAGNOSIS AND DECISION SUPPORT USING MEDICAL IMAGE ANALYSIS – CONTRIBUTIONS TO MALIGNANT MELANOMA AND COVID-19”.

Everybody is invited to attend the presentation at the Room I.0.02, or digitally via this link.

In medicine, the high volume of available data and the expanded number of treatment options have rendered it increasingly complex to determine the appropriate therapy for a specific patient. Precision medicine is a promising and emerging approach to tailor disease prevention and treatment by considering individual patient characteristics. Computer-aided diagnosis (CAD) systems can support physicians by performing fast, objective and reproducible medical image analyses and by extracting parameters that allow for more personalised disease assessment and response prediction. These features can then be used in a clinical decision support (CDS) system to guide therapeutic decisions. In this work, we investigate CAD and CDS methods for two pathologies: malignant melanoma and COVID-19.

Malignant melanoma is the most lethal form of skin cancer. Treatment planning and monitoring are generally performed using combined positron emission tomography/computed tomography (PET/CT) with fluorine-18 fluorodeoxyglucose ([18F]FDG) and regular testing of blood values. Recently, survival chances have increased due to advances in immunotherapy and targeted therapies. Nonetheless, a considerable part of this population demonstrates progressive disease. If patients with a poor prognosis can be identified before the start of therapy, a more aggressive treatment pathway could be considered to improve the survival chances.

A fully automated system was developed for lesion detection and segmentation on whole-body [18F]FDG PET/CT to extract information on the tumour load from the imaging data. We further demonstrated the feasibility of using these automatically derived imaging features in survival analysis through a comparative study with the manual method. The automated approach led to very similar results and could therefore enable the use of these parameters in clinical routine and future clinical trials.

A second pathology investigated is COVID-19, which presented great challenges for the medical sector worldwide. During the pandemic, intensive care units were overwhelmed and proper resource allocation became problematic. During the periods of high prevalence, there was an urgent need for computer-aided systems to support decisions in diagnosis, treatment and resource allocation.

In a large research collaboration, automated tools were developed to alleviate the situation. The resulting methods allow to segment lung lesions and extract relevant parameters. In addition, a model was developed to predict disease severity at one month. Its performance was validated in the context of an international challenge and proved robust through evaluation on different, multicentre datasets.

Our work demonstrated the potential of CAD and CDS systems in the field but also revealed pitfalls and shortcomings. Several challenges remain before such systems can be used readily in clinical routine, including thorough validation and medical certification. Still, important contributions were made to help in the shift towards precision medicine.

Nicolas Ospitia Patino has successfully defended his Ph.D. dissertation obtaining the greatest honors and congratulations from the jury! It’s been an incredible journey of dedication, hard work, and personal growth. He did his PhD in collaboration between MeMC and ETRO, supervised by Prof. Dimitris Aggelis (MeMC) and Prof. Johan Stiens, (ETRO), Congratulations to Nicolas for this result.

On February 1 2021 at 16.00 Tien Do Huu will defend his PhD entitled “A generic source coding. Methodology and architecture of dynamic holograms”.

Everybody is invited to attend the presentation online via https://teams.microsoft.com/l/meetup-join/19%3ameeting_N2EwMTQ2ZmMtMjA2Ny00MGQwLTk2ZDctODNmMzE3NGEyNDMz%40thread.v2/0?context=%7b%22Tid%22%3a%22695b7ca8-2da8-4545-a2da-42d03784e585%22%2c%22Oid%22%3a%2237757df7-4050-47fe-8384-aded7e9e79c3%22%7d

We address the two challenges of big heterogeneous data originated from social media and smart cities, namely data quality enhancement and data exploitation. In the first challenge, we consider several subproblems commonly found in social media data and smart city data, including user location prediction, traffic data denoising, and hyperlocal air quality prediction. For the second challenge, we aim to gain insights from data, namely we focus on detecting fake news using social media data. As there exist correlations across datapoints in social media and smart city data, we propose exploiting these correlations using graph-based deep learning techniques to address the aforementioned challenges.

Our contributions are associated with the concerned subproblems. In user location prediction, we propose a novel deep multiview model combining multiple aspects of social media data. One of the inputs of the multiview model is node representation, which is learned using a graph-deep-learning-based technique. In order to denoise traffic data, we design a special graph auto-encoder with a Kron-reduction-based pooling scheme. We devise a graph variational auto-encoder in dealing with the air quality prediction problem. For fake news detection, we propose using a graph convolutional neural network, which captures the relation between articles shared by suspicious publishers. Having experienced graph neural networks on different applications, our last contribution focuses on a more fundamental problem of regularizing graph neural networks by introducing a novel regularization technique based on dropping nodes. We conduct comprehensive experiments on benchmark datasets to verify the effectiveness of the proposed methods. Finally yet importantly, although some of our methods are designed for specific problems (e.g., air pollution prediction), our formulations are general, leading to the possibility of using these methods for different applications (e.g., recommender systems) in related domains.

AIOTI is organising a Web3 Hackathon on 21/22 September 2023 in Brussels and online.

Please feel free to register and share this information further within your communities.

About the Hackathon:

The WEB3 HACKATHON is a collaborative event with a mixed crowd of students, professionals and authorities, moderated by domain experts. The event is accessible for every individual, public or private organisations willing to enter the new digital era.

The Hackathon target is to solve a series of challenges, using WEB3 Technology:

• Industry  – Challenge: Use WEB3 to TACKLE CLIMATE EMMERGENCY
• Authorities  – Challenge: Use WEB3 to SOLVE A GOVERNANCE PROBLEM
• Media – Challenge: Use WEB3 to TACKLE FAKE NEWS
• Education – Challenge: Use WEB3 to INNOVATE EDUCATION

On November 7th 2024 at 16:00, Boris Joukovsky will defend their PhD entitled “ SIGNAL PROCESSING MEETS DEEP LEARNING: INTERPRETABLE AND EXPLAINABLE NEURAL NETWORKS FOR VIDEO ANALYSIS, SEQUENCE MODELING AND COMPRESSION”.

Everybody is invited to attend the presentation in room I.0.01 or online via this link.

There is growing use of deep learning for solving signal processing tasks, and deep neural networks (DNNs) often outperform traditional methods little domain knowledge needed. However, DNNs behave as black boxes, making it difficult to understand their decisions. The empirical approaches to design DNNs often lack theoretical guarantees and create high computational requirements, which poses risks for applications requiring trustworthy artificial intelligence (AI). This thesis addresses these issues, focusing on video processing and sequential problems across three domains: (1) efficient, model-based DNN designs, (2) generalization analysis and information-theory-driven learning, and (3) post-hoc explainability.

The first contributions consist of new deep learning models for successive frame reconstruction, foreground-background separation, and moving object detection in video. These models are based on the deep unfolding method, a hybrid approach that combines deep learning with optimization techniques, leveraging low-complexity prior knowledge of the data. The resulting networks require fewer parameters than standard DNNs. They outperform DNNs of comparable size, large semantic-based convolutional networks, as well the underlying non-learned optimization methods.

The second area focuses on the theoretical generalization of deep unfolding models. The generalization error of reweighted-RNN (the model that performs video reconstruction) is characterized using Rademacher complexity analysis. This is a first-of-its-kind result that bridges machine learning theory with deep unfolding RNNs.

Another contribution in this area aims to learn optimally compressed, quality-scalable representations of distributed signals: a scheme traditionally known as Wyner-Ziv coding (WZC). The proposed method shows that deep models can retrieve layered binning solutions akin to optimal WZC, which is promising to learn constructive coding schemes for distributed applications.

The third area introduces InteractionLIME, an algorithm to explain how deep models learn multi-view or multi-modal representations. It is the first model-agnostic explanation method design to identify the important feature pairs across inputs that affect the prediction. Experimental results demonstrate its effectiveness on contrastive vision and language models.

In conclusion, this thesis addresses important challenges in making deep learning models more interpretable, efficient, and theoretically grounded, particularly for video processing and sequential data, thereby contributing to the development of more trustworthy AI systems.

On April 25th 2025 at 16:00, Ayman Morsy will defend their PhD entitled “A NOVEL APPROACH TO DEPTH-SENSE IMAGING USING CORRELATION-ASSISTED DIRECT TIME-OF-FLIGHT”.

Everybody is invited to attend the presentation in room I.0.03 or online via this link.

Time-of-flight (ToF) imaging has emerged as a vital technology in machine vision and sensing, expanding into applications such as augmented and virtual reality, gaming, robotics, autonomous driving, autofocus, and facial recognition on smartphones and laptops. ToF technology determines the distance to an object within the detection range by emitting a light source and measuring the time it takes to return. This round-trip time determines the object’s distance, with different sensing technologies employing distinct methods to determine this time.

For ToF applications, developing sensors with high image resolution, low power consumption, and the ability to function reliably in high ambient light conditions is desirable. This dissertation presents the development of a novel single-photon avalanche diode (SPAD)-based pixel called Correlation-Assisted Direct Time-of-Flight (CA-dToF), designed for in-pixel ambient light suppression and characterized by low power consumption and a scalable pixel structure. The CA-dToF pixel uses a laser pulse correlated with two orthogonal sinusoidal signals as input to two switched capacitor channels, which average out detected ambient light while accumulating the laser pulse round-trip time.

To gain insights into CA-dToF pixel operation, both Python simulation and analytical modeling were developed. Two generations of the CA-dToF pixel were developed and characterized, with the second-generation pixel achieving the first operational performance under high ambient light conditions. The two-generation CA-dToF pixel was tested under various lighting conditions and pixel design variations. Additionally, noise sources within the pixel implementation were analyzed, and potential solutions were proposed.