Artificial Intelligence-based Innovations

FEDMix – Fusible Evolutionary Deep Neural Network Mixture Learning from Distributed Data for Robust Medical Image Analysis

Automated medical image analysis has the potential to truly innovate clinical practice by offering solutions to routine, yet key tasks, such as segmentation (i.e., delineating organs). Especially with recent advances in machine learning, in particular in deep neural networks that have led to an explosive growth of successful medical image analysis studies reported in academic literature, the time appears right for such innovations to find widespread real-world uptake. Yet, labor-intensive manual performance of these tasks is still often daily clinical practice.

In this research project, we aim to overcome a key obstacle in moving toward widespread clinical uptake of computationally intelligent medical image analysis techniques: observer variation in the definition of a ground truth.

The motivation behind this research project is to realize fundamental and clinically transferrable (in radiation oncology) innovation, validated in the real-world clinical setting of radiation treatment, by uniting and cross-fertilizing, in a novel manner, expertise in four different research fields: evolutionary algorithms (CWI/LUMC), deep (neural network) learning (eScience Center), medical image analysis (LUMC/eScience Center), Efficient Computing (eScience Center), and radiation oncology (LUMC/Amsterdam UMC, location AMC).

This research is part of the research programme Commit2Data with project number 628.011.012, which is funded by the Dutch Research Council (NWO).

DAEDALUS – Distributed and Automated Evolutionary Deep Architecture Learning with Unprecedented Scalability

To use deep learning, practitioners need to first decide what neural network architecture to use. This can be a non-trivial task, especially for new applications. Another form of AI, evolutionary algorithms, has shown much promise here, auto-designing architectures that outperform expert-designed ones, but at substantial computational overhead. In this project, we aim to improve the scalability of such neuro-evolution. We also aim to develop solutions for when data can be shared as well as when only information on learning can be shared. Finally, we focus on utilization through two applications in radiation oncology: auto-contouring of medical images and predicting dose distributions for brachytherapy cancer treatment planning.

This project involves a close collaboration with two commercial parties (Elekta and ORTEC Logiqcare) and internationally renowned experts in the field of design of evolutionary algorithms for deep neural architecture learning, analyzing and optimizing landscapes underlying classical neural networks, evolutionary algorithms, as well as experts in brachytherapy to develop and display the potential power of evolutionary machine learning to innovate brachytherapy treatment planning.

This research project (with project number 18373) is funded via the research programme Open Technology Programme by the Dutch Research Council (NWO). Part of the funding is provided by Elekta and ORTEC logiqcare.

Multi-Objective Deformable Image Registration (MODIR) – An Innovative Synergy of Multi-Objective Optimization, Machine Learning, and Biomechanical Modeling for the Registration of Medical Images with Content Mismatch and Large Deformations

By taking a multi-objective perspective on deformable image registration, we aim to remove the necessity to ad hoc fine-tune parameters, paving the road toward broad use of deformable image registration in clinical practice. In this project we further aim to solve image registration problems that are currently hardly solvable (i.e., cope with large deformations, including sliding tissue and content mismatches between image acquisitions).

LUMC and CWI collaborate closely with two commercial parties (Elekta and Xomnia) as well as end-users of the software (e.g., medical experts) to establish efficient and user-friendly ways to present, select, and use final deformable image registration outcomes. The collaboration consists of the development and validation of a novel AI-based multi-objective approaches to deformable image registration.

This research also involves collaboration with the Amsterdam UMC, location AMC.

This research is part of the research programme Open Technology Programme with project number 15586, which is financed by the Dutch Research Council (NWO), Elekta, and Xomnia. Further, the work is co-funded by the public-private partnership allowance for top consortia for knowledge and innovation (TKIs) from the Ministry of Economic Affairs.

Fast, accurate, and insightful brachytherapy treatment planning for cervical cancer through artificial intelligence

Current software for brachytherapy treatment planning requires much inefficient and not insightful human interaction, best described as improving a single treatment plan through stepwise trial-and-error. Moreover, current automation targets objectives that are not in-line with clinical evaluation criteria. This makes treatment planning highly time-consuming, very complicated, and increases the risk of treating patients with sub-optimal trade-offs between coverage of the tumor and sparing of the surrounding tissues.

We have developed a novel AI-based approach, called BRIGHT, to automated planning for brachytherapy in a prior project (collaboration between CWI and Amsterdam UMC, funded by NWO and Elekta). Different from currently available software, the underlying unique and powerful optimization engine can directly effectively target clinical evaluation criteria. The approach is furthermore more insightful, because instead of a single treatment plan, a palette of treatment plans, representing trade-offs between target coverage and organ sparing, is automatically computed. For prostate high-dose-rate brachytherapy this approach was shown to result in better treatment plans in a fraction of the time compared to current clinical practice.

For the generalization and fundamental expansion of BRIGHT to cervical cancer brachytherapy treatment, in this project we are developing and validating novel, user-friendly, and fast brachytherapy treatment planning software. The clinical translation potential will be established and demonstrated by nation-wide validation studies.

This research is a close collaboration between LUMC, CWI, and Amsterdam UMC. Further, Maastro clinic, The Netherlands Cancer Institute, Radboud UMC, UMC Utrecht, and UMC Groningen are involved in the nation-wide studies.

This research is funded by the Dutch Cancer Society (KWF Kankerbestrijding: Project N. 12183) and Elekta.

EXAMINE– Evolutionary eXplainable Artificial Medical INtelligence Engine

Machine learning (ML), especially in the form of deep learning, has shown to be capable of impressive results in many fields of science and society. Nevertheless, in most cases, learned models are still incomprehensible black boxes for humans. Especially for disciplines such as medicine, the inexplicability of ML is a bottleneck for its widespread adoption, because medical experts want to understand why a certain prediction is made.

In this project, we therefore focus on developing new forms of AI that provide models that can be insightfully inspected and understood.

We will deploy the resulting AI method, called EXAMINE (Evolutionary eXplainable Artificial Medical INtelligence Engine), on a unique and recent real-world clinical dataset of gynaecological cancer patients.

This research is a close collaboration between LUMC and CWI.  Further, the Poznan University of Technology is involved.

This research is part of the research programme Open Competition Domain Science-KLEIN with project number OCENW.KLEIN.111, which is financed by the Dutch Research Council (NWO).

TRUST-AI – Transparent, Reliable, and Unbiased Smart Tool for AI

LUMC and CWI are responsible for the health care use case of this large EU project. The focus is on the development and validation of transparent, reliable, and unbiased AI for improved clinical decision making for patients with paraganglioma.

Clinical decision making is extremely challenging for these typically slow-growing tumors that might develop progressively and therefore require specific treatment (i.e., surgery or radiation treatment) at some point in time. We are investigating if it is possible to predict the moment of onset of complaints. This would make it possible to patient-specifically adapt the follow-up with advantages for both patients and caretakers. Further, a timely chosen moment for treatment (i.e., prior to the onset of (irreversible complaints) will greatly improve the quality of life of the patients.

Further collaborations within this EU project are mainly with the University of Tartu (regarding counterfactuals) and industry partner Tazi (regarding user interface design).

This research is funded by the European Commission within the HORIZON Programme (TRUST AI Project, Contract No: 952060).

Explainable Artificial Intelligence

The focus of this project is on the development and clinical validation of explainable AI techniques for applications in medicine. For this innovation an ecosystem of close collaboration between experts with different backgrounds is needed. We aim to develop and tailor the application of such techniques for decision support in palliative oncological care and gynaecological cancer care. The ultimate goal of this project is to develop a general methodology together with guidelines so to that it can become available in an easy-to-be-tailored-to-other applications format. For this, the paraganglioma use case will be used as a test case.

This research is a close collaboration between LUMC, CWI, and Amsterdam UMC. This research is funded by the Stichting Gieskes-Strijbis Fonds.