FAUSTO MILLETARI SEGMENTATION OF MEDICAL VOLUMES USING CONVOLUTIONAL NEURAL NETWORKS
MEDICAL IMAGE SEGMENTATION ▸ Computer assisted diagnosis ▸ Computer assisted intervention ▸ Treatment Planning ▸ Enhanced visualisation
MEDICAL IMAGING ▸ Magnetic Resonance Imaging (MRI) - De-facto standard in neuroimaging - Complex and expensive - Good soft tissue contrast, limited artefacts, high SNR
MEDICAL IMAGING ▸ Ultrasound - Non-invasive, mobile, inexpensive, real-time, safe - Emits high frequency sound, forms images from tissue echoes - Noisy signal, artefacts, shadows and poor contrast - Early Parkinson’s diagnosis
A LEARNING BASED APPROACH TO SEGMENTATION ▸ Automatic segmentation ▸ Limited training data ▸ Anatomy localisation ▸ Time efficiency ▸ Prior shape knowledge
SEGMENTATION WITH CONVOLUTIONAL NEURAL NETWORKS ▸ CNNs for image segmentation - How to use CNNs for segmentation? - How to include shape priors? - What is a good network architecture choice? - Can we process 3D data? - How to be robust to limited amount of training data?
SEGMENTATION WITH CONVOLUTIONAL NEURAL NETWORKS ▸ Voxel-wise classification … CNN ▸ Segmentation mask prediction for whole volume CNN ▸ Voting strategy for localisation and segmentation x … CNN Localisation Accurate segmentation
TRAINING HOUGH-CNN ▸ Train CNN classification FG CNN Softmax BG ▸ Extract descriptors foreground patches Features (128-d) CNN f 3 f 1 f 6 f 4 f 2 f 5 Training set of (overlapping) patches and votes ▸ Build database features/votes/segmentation patches f 3 f 6 f 1 f 4 f 2 f 5
TESTING HOUGH-CNN ▸ CNN classification & Feature Extraction Features (128-d) CNN Softmax (class) Collect overlapping patches
TESTING HOUGH-CNN ▸ CNN classification & Feature Extraction Features (128-d) CNN Softmax (class) Binary classification result
TESTING HOUGH-CNN ▸ CNN classification & Feature Extraction Features (128-d) CNN Softmax (class) ▸ Use features to retrieve votes from database Find K-NN Features (128-d) Example neighbours within range r only foreground patches (circles) Binary classification result
TESTING HOUGH-CNN ▸ CNN classification & Feature Extraction Features (128-d) CNN Softmax (class) ▸ Use features to retrieve votes from database Find K-NN Features (128-d) Example neighbours within range r only foreground patches (circles) ▸ Cast votes and find peak in vote map Votes towards object centroid
TESTING HOUGH-CNN ▸ CNN classification & Feature Extraction Features (128-d) CNN Softmax (class) ▸ Use features to retrieve votes from database Find K-NN Features (128-d) Example neighbours within range r only foreground patches (circles) ▸ Cast votes and find peak in vote map Collect overlapping patches ▸ Trace back correct votes. Project associated segmentation patches (from database) ▸ ▸
EXPERIMENTAL EVALUATION Experiments Experiments Experiments • • Variation of Variation of • Variation of ▸ Ultrasound and MRI volumes – network architecture – network architecture – network architecture • different depth (3, 5 and 8 convolutional layers) • different depth (3, 5 and 8 convolutional layers) • different depth (3, 5 and 8 convolutional layers) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) – Patch size – Patch size – Patch size ▸ Network Architecture • quadratic/cubic patches with side length of 31 or 51 voxels • quadratic/cubic patches with side length of 31 or 51 voxels • quadratic/cubic patches with side length of 31 or 51 voxels – Patch dimensionality – Patch dimensionality – Patch dimensionality 2D 2D 2D 2.5D 3D 2.5D 3D 2.5D 3D ▸ Data dimensionality – training set size – training set size – training set size • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches ▸ Training dataset size • US: 1.000 or 5.000 patches/region/volume • US: 1.000 or 5.000 patches/region/volume • US: 1.000 or 5.000 patches/region/volume � ~ 80.000 or 400.000 patches � ~ 80.000 or 400.000 patches � ~ 80.000 or 400.000 patches Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks 10 10 10
EXPERIMENTAL EVALUATION ▸ Hough-CNN vs. voxel-wise classification MRI US 1 1 0,75 0,75 0,5 0,5 0,55 0,47 0,33 0,25 0,25 0,29 0 0 VOXEL-WISE CLASSIFICATION VOXEL-WISE CLASSIFICATION
EXPERIMENTAL EVALUATION ▸ Hough-CNN vs. voxel-wise classification Biggest Training set Smallest Training Set
EXPERIMENTAL EVALUATION ▸ Hough-CNN vs. voxel-wise classification
EXPERIMENTAL EVALUATION ▸ Patch dimensionality (2D, 2.5D, 3D) Experiments Experiments Experiments Experiments Experiments Experiments 1 • Variation of • Variation of • Variation of • Variation of • Variation of • Variation of 0,85 0,83 0,82 – network architecture – network architecture – network architecture – network architecture – network architecture – network architecture 0,75 0,77 • different depth (3, 5 and 8 convolutional layers) • different depth (3, 5 and 8 convolutional layers) • different depth (3, 5 and 8 convolutional layers) • different depth (3, 5 and 8 convolutional layers) • different depth (3, 5 and 8 convolutional layers) • different depth (3, 5 and 8 convolutional layers) 0,68 0,68 • different convolution filter sizes (side length of 3,5,7,9,11 voxels) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) • different convolution filter sizes (side length of 3,5,7,9,11 voxels) – Patch size – Patch size – Patch size – Patch size – Patch size – Patch size 0,5 • quadratic/cubic patches with side length of 31 or 51 voxels • quadratic/cubic patches with side length of 31 or 51 voxels • quadratic/cubic patches with side length of 31 or 51 voxels • quadratic/cubic patches with side length of 31 or 51 voxels • quadratic/cubic patches with side length of 31 or 51 voxels • quadratic/cubic patches with side length of 31 or 51 voxels – Patch dimensionality – Patch dimensionality – Patch dimensionality – Patch dimensionality – Patch dimensionality – Patch dimensionality 2D 2.5D 3D 2D 2.5D 3D 2D 2.5D 3D 2D 2.5D 3D 2D 2.5D 3D 2D 2.5D 3D 0,25 2D 2.5D 3D 2D 2.5D 3D – training set size – training set size – training set size – training set size – training set size – training set size 0 • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches • MRI: 100,1.000 or 10.000 patches/region/volume � ~ 116 . 000, 765 . 000 or 3 . 100 . 000 patches MRI US • US: 1.000 or 5.000 patches/region/volume • US: 1.000 or 5.000 patches/region/volume � ~ 80.000 or 400.000 patches � ~ 80.000 or 400.000 patches • US: 1.000 or 5.000 patches/region/volume • US: 1.000 or 5.000 patches/region/volume � ~ 80.000 or 400.000 patches � ~ 80.000 or 400.000 patches • US: 1.000 or 5.000 patches/region/volume • US: 1.000 or 5.000 patches/region/volume � ~ 80.000 or 400.000 patches � ~ 80.000 or 400.000 patches 2D Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks 10 10 Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks 10 10 Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks Segmentation of MRI and Ultrasound Scans Using Deep Convolutional Neural Networks 10 10
EXPERIMENTAL EVALUATION ▸ Network architectures MRI US 1 1 0,85 0,85 0,84 0,83 0,83 0,83 0,82 0,75 0,75 0,77 0,77 0,72 0,72 0,71 0,7 0,68 0,5 0,5 0,25 0,25 0 0 3-3-3-3-3 3-3-3-3-3-3-3-3 5-5-5-5-5 7-5-3 9-7-5-3-3 SMALL ALEX 3-3-3-3-3 3-3-3-3-3-3-3-3 5-5-5-5-5 7-5-3 9-7-5-3-3 SMALL ALEX Architecture Architecture 2D
Recommend
More recommend
