PET image classification using HHT-based features through fractal sampling

PET image classification using HHT-based features through fractal sampling 7th. International Work-Conference on the Interplay between Natural and Artificial Computation June 20, 2017 Andrés Ortiz1Francisco Lozano1Alberto Peinado1 María J. García-Tarifa3Juan M. Górriz2Javier Ramírez2 1Department of Communications Engineering Universidad de Málaga, Spain. 2Dept. of Signal Theory, Networking and Communications Universidad de Granada, Spain 3Dept. Pharmacology and Pediatrics, Universidad de Málaga, Spain 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology Image Preprocessing Brain parcellation Sampling Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Presentation Outline Introduction Methodology Image Preprocessing Brain parcellation Sampling Feature Extraction Results Conclusions and Future Work 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. 2Introduction Methodology Image Preprocessing Brain parcellation Sampling Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Introduction 1 IAlzheimer’s disease (AD) is the first most common neurodegenerative disorder in the elderly. Currently, there is no cure for AD and their causes remain unknown. A precise diagnosis plays a decisive role to start the treatment in the early stages of the disease. However, it remains a challenge →CAD systems IA number of neuroimage modalities have been proposed for its use in the differential diagnosis of AD: IMagnetic Resonance Imaging (MRI) ISingle Photon Emission Computerized Tomography (11C-SPECT) IPositron Emission Tomography (18F-FDG PET) ISeveral works using statistical and machine learning techniques have been proposed to extract relevant patterns in the images: IMultivariate methods (PCA, ICA, etc.) IImage analysis techniques (texture analysis). 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. 3Introduction Methodology Image Preprocessing Brain parcellation Sampling Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Introduction 1 IIn this work we used a different approach: IObjective: use 1D signal processing methods to extract relevant features from 3D images. IChallenge: Transforming a 3D images to 1D signals while preserving 3D neighbourhood (voxel spatial relationship) IDatabase: 68 CN subjects + 70 AD patients 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology 4Image Preprocessing Brain parcellation Sampling Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Methodology Image preprocessing IPET images from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. ICo-registration: each image is spatially normalized to the MNI space (PET Template) IIntensity normalization was applied to be able to compare the uptake value in areas of specific activity. 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology Image Preprocessing 5Brain parcellation Sampling Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Methodology Brain Parcellation Brain is parcelled into 116 regions according to the Automated Anatomical Labelling (AAL) atlas The best-fitting parallelepiped is computed for each region Only the most 42 relevant regions are extracted, according to Huang et. al, 2009 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology Image Preprocessing Brain parcellation 6Sampling Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Methodology Sampling IOnce the images have been normalized, voxels are sampled throughout the 3D volume: 1. The 3D image is converted into a sequence of intensity values 2. Spatial neighbourhood has to be preserved, keeping the relationship between voxels, as this is an essential source of information. 3. Neighbour voxels in the 3D space should be also neighbours in the 1D space. ISampling is performed by means of Hilbert-Peano 3D homogeneous fractal curves. IBasically, it is a function f:R→Rn IContinuity is preserved →adjacency condition IThe curve is uniquely defined by fixing initial and final subintervals and the rotation matrix IIt can be generated by the iterative application of affine transformations 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology Image Preprocessing Brain parcellation 7Sampling Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Methodology Sampling(2) −1 0 1 2 3 4 5 6 7 8 −1 0 1 2 3 4 5 6 7 8 x y (a) 2 4 6 8 10 2 4 6 8 10 3 4 5 6 7 8 9 10 x y z (b) Figure: Example of 2D (a) and 3D (b) Hilbert curves 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology Image Preprocessing Brain parcellation Sampling 8Feature Extraction Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Methodology Feature extraction. Empirical Mode Decomposition (EMD) IEMD allows decomposing a signal into AM and FM components, namely Intrinsic Mode Functions (IMF) along with a trend component (residue). IThe main advantage of EMD is that can be applied to non stationary and non-linear signals. IUnlike other decomposition methods such as Fourier decomposition or Wavelet decomposition, EMD does not use predefined basis functions. The basis is empirically computed by the so called sifting method. IThe Sifting process consists on: 1. Identify all the local extrema in the signal. 2. Connect all the local maxima by a cubic spline line as the upper envelope. 3. Repeat the procedure for the local minima to produce the lower envelope. 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology Image Preprocessing Brain parcellation Sampling Feature Extraction 15 Results Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Results ROC analysis IROC curve for the classifier reports an AUC of 0.95. 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 False Positive Rate (1−Specificity) True Positive Rate (Sensitivity) HHT−method PCA VAF Cut−off point Random Classifier 16 PET image classification using HHT-based features through fractal sampling A. Ortiz et al. Introduction Methodology Image Preprocessing Brain parcellation Sampling Feature Extraction Results 16 Conclusions and Future Work Department of Communications Engineering Universidad de Málaga Conclusions and Future Work Conclusions IA 3D fractal sampling method is used over brain regions to convert 3D images into time-varying signals. ITime series analysis techniques can be used. IWe proposed the use of Hilbert-Huang transform to extract features IDiscriminative features are computed providing a classification accuracy values up to 92% and AUC of 0.95 outperforming the VAF approach and PCA approaches. Future work IUse the same method to implement Functional PCA based on EMD basis IFPCA EMD based to model the voxel intensity variations in time →longitudinal analysis Thank you for your attention! Questions? Comments? Volunteers? aor[email protected]