Deep learning for neuroimaging-based disease decoding

• Applicant: Professorin Dr. Kerstin Ritter
• Subject Area: Clinical Neurology; Neurosurgery and Neuroradiology
• Term: from 2017 to 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 389563835

Recently, deep learning approaches (in particular convolutional networks) have been extremely successful in solving supervised learning problems such as pattern recognition. In this research proposal, we suggest using convolutional networks for finding new representations from neuroimaging data in order to predict disease conversion and future clinical disability. In particular, we will investigate the conversion from (1) clinically isolated syndrome to clinically definite multiple sclerosis and (2) mild-cognitive impairment to Alzheimer's disease. Additionally, we will predict in both cases disease progression in terms of clinical and neuropsychological disability. The analysis will be based on three MRI sequences, namely MPRAGE, FLAIR and DTI. Whereas previous disease decoding approaches mostly relied on expert-based extraction of features in combination with standard classification algorithms and thus strongly depend on the choice of data representation, convolutional networks are capable of learning hierarchical information directly from raw imaging data. By this, they have a great potential for finding unexpected and latent data characteristics and might perform as a real "second reader". Since the clinical interpretation is of major importance in disease diagnostics and convolutional networks are usually considered as "black box" models, we will use here a a state-of-the-art visualization technique called layer-wise relevance propagation to make the learned content of convolutional neural networks visible. Resulting feature maps will be discussed with two radiologists regarding their clinical relevance for disease prediction. To further improve prediction accuracy, clinical and neuropsychological scores will be taken into account in the framework of multimodal data analysis. We think that deep learning-based MRI representations will not only lead to a more accurate disease prediction and characterization of disease course in both Alzheimer's disease and multiple sclerosis but also to a better understanding of MRI features that contribute to the one or the other diagnosis in single subjects.

• DFG Programme: Research Grants
• Co-Investigators: Professor Dr. John-Dylan Haynes; Professor Dr. Friedemann Paul