MIAAI's aim is to explore applications in all areas of medicine in collaboration with the group's affiliated clinical scientists, and then to develop advanced methods tailored to each speciality.
​
Our main areas of focus are detailed below. Meet our Core Team and affiliated Clinical Scientists or get in touch to talk to a member of our team.
MEDICAL IMAGING
Medical imaging comprises a variety of different modalities such as mammography, computed tomography (CT) and magnetic resonance imaging (MRI), and is used for diagnostic purposes, as the basis for treatment decisions, for planning and performing medical interventions such as surgery and radiotherapy and for research purposes.
Medical image analysis involves the study of 2D and 3D image data sets obtained from the human body using the above modalities. A prominent feature of medical imaging is that it provides a detailed but non-invasive examination of human anatomy and physiology and reveals pathological changes. Imaging is thus an integral part of medical diagnostics and essential for therapy planning in a wide range of diseases.
For the routine clinical evaluation and reporting of medical imaging, usually only individual quantitative measurements (e.g. tumour diameter in one or two planes) and a qualitative image assessment are recorded. However, detailed assessments of the image content (such as the exact volume of a tumour and its appearance on the different imaging modalities) could provide important information about the disease. With the current technology, such detailed quantitative evaluation is very time-consuming and therefore not integrated into routine clinical care.
However, systematic quantitative image analysis would enable researchers to capture relevant disease characteristics that could be integrated into intelligent clinical decision-making systems, and thus simplify and improve diagnostics and treatment planning.
RADIOMICS
& QUANTITATIVE IMAGE ANALYSIS
Radiomics refers to a type of analysis of medical images in which a large number of quantitative image features – of a tumour, for instance – are extracted using computer algorithms. Radiomics make it possible to capture the appearance of tumours on images quantitatively and in greater detail than in routine radiological reporting.
In a radiomics analysis pipeline, the first step of image processing is often the outlining of the diseased structures, such as a tumour, on the medical images (= segmentation; creation of regions of interest, ROIs).
Afterwards, specially developed computer programs allow the calculation of radiomics features, which primarily describe tumour size, shape and texture with the help of numbers. Statistical methods and machine learning techniques are then used to select the most meaningful features and to make clinically meaningful predictions based on the image material (e.g. treatment response, prognosis, tumour subtype, tumour-specific gene mutations, etc.).
IMAGING
(A)
SEGMENTATION
(B)
RADIOMICS
Tumour Size & Shape
Volume
Diameter along different​
​axes
Compactness
Frequency & spatial distribution of grey levels
Texture Analysis
GLCM (Grey level co-occurrence matrix)
Histogram Analysis
Skewness
Contrast=
Entropy=
(C)
AI
for the prediction of Clinically
meaningful variables
Survival
Tumour Size
(D)
Illustration of a radiomics pipeline using breast cancer imaging as an example. After the medical images are taken (a), the tumour is manually marked on MRI images (segmentation = creation of regions of interest, ROIs; b). Radiomics features are then calculated according to predefined formulas (c): for example, the tumour volume and various shape features such as diameter along different axes, compactness, sphericity (similarity to spherical shape) are calculated. Based on the pixels and their grey levels within the tumour, histogram analysis parameters are calculated based on their frequency distribution, and texture features are calculated considering the spatial distribution and neighbourhoods of the pixels (such as the features of the grey-level co-occurrence matrix). Finally, statistical methods and AI techniques are used to reduce the number of relevant radiomics features and develop predictors from the remaining radiomics for predicting, for example, patient prognosis or treatment response.
ARTIFICIAL INTELLIGENCE
IN MEDICAL IMAGING
Artificial intelligence (AI) is an emerging field of research that includes both classical machine learning and deep learning (DL).
Computer algorithms inspired by the functioning of human intelligence, such as complex neural networks, are used to analyse complicated data.
Medical imaging is a very promising clinical application of AI. Classical machine learning techniques are often used together with the computation of predefined radiomics features.
However, deep learning using deep neural networks is increasingly being used in automated image analysis. This usually requires larger datasets, but here it automatically extracts features from complex medical images instead of predefined radiomics features, and often yields gains in diagnostic accuracy.
Although DL methods are computationally intensive, they can be accelerated by high-performance computing and parallel computing techniques.
GROUP ACTIVITIES
MEDICAL IMAGE ANALYSIS & ARTIFICIAL INTELLIGENCE
The Medical Image Analysis and Artificial Intelligence (MIAAI) group at the DPU focuses on the analysis of medical imaging for the development of quantitative biomarkers, as well as the use of AI to predict the presence of disease, its progression or treatment response. Currently, it is focusing on cancer and the prediction of prognosis and treatment response, but the group’s work can be applied to any area of medical imaging due to its versatility.
Manual labelling of pathological changes on medical images (segmentation) is essential for many quantitative analyses as well as radiomics. This is usually performed by one or more specialised radiologists/clinicians, requires specialised knowledge and is very time-consuming. Although this type of segmentation is the gold standard, significant variability between different performing specialists has been demonstrated. To reduce time and variability, the development of automated segmentation methods using AI will also be a focus of the MIAAI research group. Fully automated segmentation and image analysis algorithms are a key element of smart medicine and are essential for standardising the methods mentioned.
Radiomics features can be affected by variability of segmentations and different scanning techniques and image reconstruction methods. The lack of standardisation of radiomics features in different imaging modalities (MRI, CT, mammography, etc.) remains an unsolved problem. The MIAAI group is investigating the use of 3D printed phantoms to explore radiomics features and their robustness for different clinical applications and settings. The reproducibility and repeatability of many radiomics features are investigated using phantom and patient data to develop more robust radiomics predictors in the future.
HEALTHCARE EVALUATION
OF THE EXPECTATIONS, FEARS AND PERCEPTIONS OF DIFFERENT PLAYERS
Another focus of the research group is to evaluate positive and negative expectations among patients, physicians and other healthcare providers from different disciplines towards AI and its integration into the healthcare system. The knowledge of different actors in the healthcare system with regard to AI will be surveyed, investigating whether and how knowledge transfer could improve acceptance of AI systems.
The results of these questionnaire-based surveys and interviews will be used to define action programmes and objectives that will enhance knowledge and manage expectations in the general population as well as among healthcare providers. A long-term goal is to establish a network including patients, clinicians and researchers.
Within this network, benchmark workshops held in cooperation with clinics and healthcare providers will support AI knowledge transfer and the promotion of new evidence-based solutions. Through these measures, we aim to identify essential barriers and facilitators of AI implementation and increase both individual and organizational health literacy pertaining to AI.
The evaluation may also provide impulses and stimuli for politicians and professional associations in shaping their policies regarding the implementation of new systems.