Deep Feature Learning and Cascaded Classifier for Large Scale Data
Research output: Book/Report › Ph.D. thesis › Research
Standard
Deep Feature Learning and Cascaded Classifier for Large Scale Data. / Prasoon, Adhish.
Department of Computer Science, Faculty of Science, University of Copenhagen, 2014. 117 p.Research output: Book/Report › Ph.D. thesis › Research
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - BOOK
T1 - Deep Feature Learning and Cascaded Classifier for Large Scale Data
AU - Prasoon, Adhish
PY - 2014
Y1 - 2014
N2 - This thesis focuses on voxel/pixel classification based approaches for image segmentation. The main application is segmentation of articular cartilagein knee MRIs. The first major contribution of the thesis deals with largescale machine learning problems. Many medical imaging problems need hugeamount of training data to cover sufficient biological variability. Learningmethods scaling badly with number of training data points cannot be used insuch scenarios. This may restrict the usage of many powerful classifiers havingexcellent generalization ability. We propose a cascaded classifier whichallows usage of such classifiers in large scale problems. We demonstrate itsapplication for segmenting tibial articular cartilage in knee MRI scans, withnumber of training voxels being more than 2 million. In the next phaseof the study we apply the cascaded classifier to a similar but even morechallenging problem of segmenting femoral cartilage. We discuss similaritiesand provide our solutions to the challenges. Our cascaded classifier forcartilage segmentation comprised of two stages of classification combiningnearest neighbour classifier and support vector machine. We compared ourmethod to a state-of-the-art method for cartilage segmentation using onestage nearest neighbour classifier. Our method achieved better results thanthe state-of-the-art method for tibial as well as femoral cartilage segmentation.The next main contribution of the thesis deals with learning featuresautonomously from data rather than having a predefined feature set. Weexplore deep learning approach of convolutional neural network (CNN) forsegmenting three dimensional medical images. We propose a novel systemintegrating three 2D CNNs, which have a one-to-one association with thexy, yz and zx planes of 3D image, respectively and this system is referredas triplanar convolutional neural network in the thesis. We applied thetriplanar CNN for segmenting articular cartilage in knee MRI and comparedits performance with the same state-of-the-art method which was used asa benchmark for cascaded classifier. Although our method used only 2Dfeatures at a single scale, it performs better than the state-of-the-art methodusing 3D multi-scale features. In the latter approach, the features and theclassifier have been carefully adapted to the problem at hand. That we wereable to get better results by a deep learning architecture that autonomouslylearns the features from the images is the main insight of this study.While training the convolutional neural networks for segmentation purposes,the commonly used cost function does not consider the labels of theneighbourhood pixels/voxels. We propose spatially contextualized convolutionalneural network (SCCNN) which incorporates the labels of the neighbouringpixels/voxels while training the network. We demonstrate its applicationfor the 2D problem of segmenting horses from the Weizmann horsesdatabase using 2D CNN and our 3D problem of segmenting tibial cartilagein knee MRIs using triplanar CNN. The proposed SCCNN improved thesegmentation performance in both the cases. The good results obtained bySCCNN encourage to gain more insight into such frameworks.
AB - This thesis focuses on voxel/pixel classification based approaches for image segmentation. The main application is segmentation of articular cartilagein knee MRIs. The first major contribution of the thesis deals with largescale machine learning problems. Many medical imaging problems need hugeamount of training data to cover sufficient biological variability. Learningmethods scaling badly with number of training data points cannot be used insuch scenarios. This may restrict the usage of many powerful classifiers havingexcellent generalization ability. We propose a cascaded classifier whichallows usage of such classifiers in large scale problems. We demonstrate itsapplication for segmenting tibial articular cartilage in knee MRI scans, withnumber of training voxels being more than 2 million. In the next phaseof the study we apply the cascaded classifier to a similar but even morechallenging problem of segmenting femoral cartilage. We discuss similaritiesand provide our solutions to the challenges. Our cascaded classifier forcartilage segmentation comprised of two stages of classification combiningnearest neighbour classifier and support vector machine. We compared ourmethod to a state-of-the-art method for cartilage segmentation using onestage nearest neighbour classifier. Our method achieved better results thanthe state-of-the-art method for tibial as well as femoral cartilage segmentation.The next main contribution of the thesis deals with learning featuresautonomously from data rather than having a predefined feature set. Weexplore deep learning approach of convolutional neural network (CNN) forsegmenting three dimensional medical images. We propose a novel systemintegrating three 2D CNNs, which have a one-to-one association with thexy, yz and zx planes of 3D image, respectively and this system is referredas triplanar convolutional neural network in the thesis. We applied thetriplanar CNN for segmenting articular cartilage in knee MRI and comparedits performance with the same state-of-the-art method which was used asa benchmark for cascaded classifier. Although our method used only 2Dfeatures at a single scale, it performs better than the state-of-the-art methodusing 3D multi-scale features. In the latter approach, the features and theclassifier have been carefully adapted to the problem at hand. That we wereable to get better results by a deep learning architecture that autonomouslylearns the features from the images is the main insight of this study.While training the convolutional neural networks for segmentation purposes,the commonly used cost function does not consider the labels of theneighbourhood pixels/voxels. We propose spatially contextualized convolutionalneural network (SCCNN) which incorporates the labels of the neighbouringpixels/voxels while training the network. We demonstrate its applicationfor the 2D problem of segmenting horses from the Weizmann horsesdatabase using 2D CNN and our 3D problem of segmenting tibial cartilagein knee MRIs using triplanar CNN. The proposed SCCNN improved thesegmentation performance in both the cases. The good results obtained bySCCNN encourage to gain more insight into such frameworks.
UR - https://soeg.kb.dk/permalink/45KBDK_KGL/fbp0ps/alma99122494029405763
M3 - Ph.D. thesis
BT - Deep Feature Learning and Cascaded Classifier for Large Scale Data
PB - Department of Computer Science, Faculty of Science, University of Copenhagen
ER -
ID: 113690033