Biomedical Imaging and Genetic (BIG) Data Analytics in Dementia and Oncology “Deep Learning for Precision Medicine” Wiro Niessen Biomedical Imaging Group Rotterdam Departments of Radiology & Medical Informatics Erasmus MC Imaging Physics Faculty of Applied Sciences, Delft University of Technology Quantib (disclosure) Precision medicine: what do we want? Taking individual variability into account to optimize diagnosis, prognosis and treatment Precision medicine: what does it mean? • Cardiovascular disease: knowing who to treat • Cancer treatment: predict what treatment is likely successful • Dementia: prognosis to support preventive strategies 1
Precision medicine: what does it mean? • Cardiovascular disease: knowing who to treat • Cancer treatment : predict what treatment is likely successful • Dementia: prognosis to support preventive strategies 1967 – 34 years 66 years 63 years 64 years 2000 – 67 years 65 years William Utermohlen (1933-2007) Possible early markers for dementia Jack et al., 2010; 2013 2
Rotterdam Study Population Study initiated in 1990 healthy diagnosis screening etiology prediction diseased population-based studies clinical studies Population imaging study Brain Risk factors: Outcome: black box Atrophy, lesions Genetic Dementia Stroke Lifestyle Smoking 3
Can this approach succeed? De Gro root et al al, “St “Stro roke 2013 2013 Pro rogre ress ss and and Inno nnovat ation n aw award ard” Conclusion What we currently see visually (as appreciable white matter lesions) is only the tip of the iceberg of white matter pathology: searching for QIBs logical next step 12 4
Rotterdam Scan Study (> 14.000 MRI data acquired) • Tissue quantification • Lesion assessment • Segmentation & shape • Microstructure & function • Incidental brain findings • Cerebral microbleeds Library of quantitative imaging biomarkers Brain structures Hippocampal shape and volume Subcortical WML White matter tracts Structural connectivity Role deep learning § Quantitative imaging biomarker (QIB) extraction is increasingly replaced with deep learning § Deep learning promising for extracting QIB that hitherto were difficult to detect/extract. 5
Enlarged Perivascular Spaces (EPVS) § EPVS: emerging biomarker for AD, Stroke, MS § Space between blood vessels and pia mater § Filled with interstitial and cerebrospinal fluid § T2/PD § Features § Tubular § Hyper-intense § Challenges § Difficult to quantify EPVS (blue) in the brain in axial plane. PD. Rotterdam Study Database. § Variable shape EPVS - Ground Truth annotation. § Limit of the resolution Rotterdam Study Database. Data - Enlarged Perivascular Spaces § Ground truths in 4 brain regions for 2000 to 5000 participants § 2 kinds of Ground truths: visual scoring (number), dot annotations 24 3 Method: GP-Unet Detection with Weak Labels ¹ !. Dubost, F., Bortsova, G., Adams, H., Ikram, A., Niessen, W., Vernooij, M. and De Bruijne, M., 2017. GP-Unet: Lesion Detection from Weak Labels with a 3D Regression Network.MICCAI 2017. 6
Correlation with visual scoring Region Interrater Intrarater Correlation Automatic/ Visual Basal Ganglia 0.62 0.80 0.801 +/- 0.026 CSO 0.80 0.88 0.872 +/- 0.024 Hippocampus 0.82 0.85 0.819 +/- 0.030 Midbrain 0.75 0.82 0.622 +/- 0.042 CADDementia Challenge Computer-aided diagnosis of dementia based on structural MRI Example data Healthy MCI Alzheimer Oncology: Concept of Radiomics Radiomics Hypothesis : There exists a correlation between medical image features and underlying biological information. Image adopted from Lambin et al. 2012 7
Radiomics Platform 1p/19q Mutation in Low Grade Glioma Goal Predict 1p/19q genetic mutations in 81 patients Ground Truth PA Modality MR (T2, DWI) Discovery Two-class: co-deleted vs non co-deleted Deep learning in neuro oncology Measure AUC [0.68; 0.87] Classify tumor subtype: F1-Score [0.81; 0.95] Sensitivity [0.70; 1.00] Specificity [0.80; 0.99] Tumor segmentation Location-specific mutations: guide biopsies (pre-processing step) 8
Population Imaging Genetics Integrating imaging and genetics for improved understanding of disease processes, and improved detection, diagnosis and therapy planning & guidance Predicting is not easy … AD risk genes Future of imaging genetics § Holy grail: Find phenotype = f (genotype, environmental factors) § Current approaches: mostly massive number of linear regressions § Promises in: § Larger datasets § Machine and deep learning for learning more complex relations § Challenges § DL and ML cannot straightforwardly be applied (heterogeneous data, biological variability) § Modular approach, integrating prior knowledge with DL 9
Deep learning in population imaging genetics § Increasingly the state of the art in QIB extraction § Still challenging to solve “clinically relevant” applications § Complex, heterogeneous data; few training samples § Defining right scope § How to integrate prior knowledge § Powerful tool to support investigating challenging relation between genetic data, environmental factors and phenotype Acknowledgements Erasmus MC Department of Radiology/Neuroradiology Bas Jasperse, Marion Smits, Aad van der Lugt, Meike Vernooij, Tom den Heijer Department of Epidemiology: Monique Breteler, Arfan Ikram, Marielle Poels, Ben Verhaaren Biomedical Imaging Group Rotterdam: Hakim Achterberg, Renske de Boer, Esther Bron, Marleen de Bruijne, Florian Dubost, Marius de Groot, Wyke Huizinga, Stefan Klein, Marcel Koek, Carolyn Langen, Fedde van der Lijn, Dirk Poot, Genady Roshchupkin, Erwin Vast, Henri Vrooman, Sebastiaan van der Voort, Martijn Starmans TU Delft: Frans Vos, Lucas van Vliet Imperial College London Alexander Hammers, Daniel Rueckert 10
Recommend
More recommend
