Projects / Programmes
Machine learning based medical image analysis for prognosis of brain disease course and therapy efficacy
| Code |
Science |
Field |
Subfield |
| 2.06.10 |
Engineering sciences and technologies |
Systems and cybernetics |
Medical informatics |
| Code |
Science |
Field |
| 2.06 |
Engineering and Technology |
Medical engineering
|
image analysis, image processing, machine learning, predictive models, magnetic resonance imaging, neurodegenerative diseases, multiple sclerosis, disease course prediction, therapy efficacy assessment
Researchers (31)
Organisations (2)
Abstract
This project will directly address the priority field Research on neurodegenerative diseases (JPND) as set by the Slovenian research agency. Without loss of generality, this project will focus on multiple sclerosis (MS), since it involves two common pathologies also found in cerebrovascular disease, Alzheimer's and other dementias, depression, schizophrenia and bipolar disorder, etc. The two pathologies of interest are scar tissue or lesions and progressive loss of neurons or neurodegeneration, which typically are one of the few paraclinical symptoms that appear from several months up to a few years before any clinical symptoms can even be recognized. Magnetic resonance (MR) tomographic imaging is by far the most sensitive soft tissue imaging technique and therefore extensively used for the assessment of normal brain structure status and the detection of pathologic lesions. Based on brain MR scans the lesion accumulation and degree of neurodegeneration may be quantified in vivo. This is also the reason why information-rich brain MR images are being increasingly utilized in large-scale clinical trials, for instance, for testing new drugs and therapies the MR image based measures are already used as surrogates of clinical outcome measures. In this way, drug development proceeds faster and is thus less hazardous for the enrolled patients. Whereas clinical trials involve between-group comparisons, clinical management of individual MS patients is more challenging and the use of MR image based measures or imaging biomarkers has not yet proliferated. Main difficulties are high inter- and intra-observer variability of visual assessment of the MR images and high variability in quality of standard-of-care MR images. The emerging technical solution leading to more objective and reliable quantitative assessment is a computational analysis of MR images. In this project, we will leverage advanced machine learning tools for MR image analysis with the goal to develop accurate, interpretable and robust prediction models that address the need for personalized early prognosis of disease course and treatment efficacy for MS and other neurodegenerative diseases from standard-of-care MR images. The proposed project has 8 deliverables, which involve (1) acquisition and annotation of brain and spinal cord MR images of MS patients and (2) collection of associated set of clinical, laboratory, gait, balance and self-reported outcome measures. For the standard-of-care MR images (3) automated pipeline including MR sequence identification and image quality assessment and (4) adversarial based image enhancement will be developed, followed by (5) feature-rich MR image description based on brain and spinal cord segmentation, radiomics and shape features, and autoencoder representations. Next, (6) based on features and enhanced MR images we will develop novel and improved prediction models, focusing also on interpretability and robustness. Besides (7) expected impactful research publications, the ultimate goal is to (8) develop, integrate and prospectively validate a decision support system for managing MS based on the state-of-the-art prediction models and verify its capabilities using routine, standard-of-care MR images. Preliminary results suggest that MR image based treatment efficacy prediction enables the clinicians to determine the optimal treatment within 1.2 years on average, compared to current average of 3.9 years based on using clinical outcome measures like the extended disability status score (EDSS) or relapse occurrence. Besides, through the search for the optimal prediction models we also aim to elucidate the rather notorious associations between the MR imaging data and clinical outcomes in individual MS patients. The proposed methods and systems have a clear benefit for the patient and clinical MS disease management, but may also help decrease the high socioeconomic costs associated with the MS disease.
