IPSJ SIG Technical Report Vol.2015-MPS-103 No.29 Vol.2015-BIO-42 No /6/24 Deep Convolutional Neural Network 1,a) 1,b),c) X CT (Computer Aided D

Size: px

Start display at page:

Download "IPSJ SIG Technical Report Vol.2015-MPS-103 No.29 Vol.2015-BIO-42 No /6/24 Deep Convolutional Neural Network 1,a) 1,b),c) X CT (Computer Aided D"

しょうぶすみい
5 years ago
Views:

1 Deep Convolutional Neural Network 1,a) 1,b),c) X CT (Computer Aided Diagnosis : CAD) CAD Deep Convolutional Neural Network (DCNN) DCNN CT DCNN DCNN Support Vector Machine DCNN, Anaysis for Deep Convolutional Neural Network feature with Diffuse Lung Disease classification Satoshi Suzuki 1,a) Hayaru Shouno 1,b) Shoji Kido,c) Abstract: The computer aided diagnosis (CAD) system is desired to develop for supporing physicians to diagnose the diffuse lung diseeases (DLD). We apply a deep convolutional nerual network (DCNN) into the CAD system for the classification of DLDs. DCNN is a kind of multi layer neuralnetwork which can automatically extract features description depending on the input data. Meanwhile, obtaining the effective DCNN features, it requires large amount of training data. In the field of medical image analysis, the number of acquired data sometimes is not enough to train. Thus, in our previous work, we introduce a kind of semi-supervised learning (SSL) based method for training the DCNN. At first, we apply massive natural images, which we can easily collect, for the unlabeled data. After that, we train the pre-trained DCNN with the small number of the DLD patterns as the labeled data. In this research, we analyze the intermediate layers of the DCNN in order to investigate the feature expression transition by use of the linear support vector machine (SVM). Keywords: Deep Convolutional Neural Network Semi-supervised Learning, Diffuse Lung Disease image analysis 1, a) s @edu.cc.uec.ac.jp b) shouno@uec.ac.jp c) kido@ai.csse.yamaguchi-u.ac.jp 1. 1 CT ( ) 1

i p convolution k x fi conv (k, x) : f conv p,i (k, x) = l,u g i (k, l, u) f pool p,i 1 (l, x u), (1) 1 CT Deep Convolutional Neural Network (DCNN) (Computer Aided Diagnosis: CAD) DCNN Fukushima

2 i p convolution k x fi conv (k, x) : f conv p,i (k, x) = l,u g i (k, l, u) f pool p,i 1 (l, x u), (1) 1 CT Deep Convolutional Neural Network (DCNN) (Computer Aided Diagnosis: CAD) DCNN Fukushima Neocognitron [1][2][3] CT [4] DCNN DCNN 2. Deep Convolutional Neural Network (DCNN) Fukushima Neocognitron DCNN [5] 2.1 DCNN 10 DCNN 2 Krizhevsky O(10 7 ) [6] DCNN 2 convolution ReLU pooling normalize Neocognitron (stage) [1] g i (k, l, u) ( i 1 ) pooling l i convolution k ReLU convolution Rectified Linear Unit (ReLU: ) ReLU fp,i ReLU (k, x) = max[0, fp,i conv (k, x)] (2) pooling max-pooling i k x max-pooling f pool i (k, x) f pool p,i (k, x) = max [ f p,i ReLU (k, u)], (3) u N(x) [7] N(x) x pooling pooling k, x : f pool p,i = { f pool p,i (k, x)} k,x (4) 2 Krizhevsky DCNN fc6 DCNN softmax (Back-propagation: BP) soft max f k p(c k f) p(c k f) = p(x C k)p(c k ) j p( f C j )p(c j ) = exp(a k) j exp(a j ) a k a k = ln(p( f C k )p(c k )) [8] 2.2 DCNN DCNN LeCun (5) 2

3 Pooling Convolutions Normalize Convolutions Pooling (a) (b) (c) (d) Consolidation GGO Honeycomb Reticular IIPs (e) (f) (g) Emphysema Nodular Normal Convolutions Data conv1 ReLU1 pool1 Normalize norm1 conv2 ReLU2 Pooling Convolutions Convolutions conv3 conv4 pool2 norm2 ReLU3 ReLU4 conv5 relu5 pool5 Fully-Connect fc6 fc7 relu6 relu7 fc8 drop6 drop7 loss 3 ROI 7 a b c d i-th Stage connections l x k x x LOPOCV Pool i-1 conv i ReLU i Pool i 2 DCNN ( ) ( ) DCNN 10 3 O(10 7 ) Rectified Linear Unit (ReLU) DCNN 6 DCNN O(10 4 ) MNIST Le 9 DCNN O(10 8 ) YouTube [3] Krizhevsky DCNN O(10 4 ) CT DCNN O(10 6 ) [4] DCNN Support Vector Machine (SVM) [9] Leave-One-Person-Out CV (LOPOCV) (Leave One Person Out Cross Varidation: LOPOCV) X CT Leave-One-Out CV 3.2 X CT 7 4 (consolidation: CON) (ground-glass opacity: GGO) (honeycomb: HCM) (reticular: RET) 2 (emphysema: EMP) (nodular: NOD) (Normal: NOR) 7 CT [pixel] ROI (Region of Interest) [pixel] 80 3 CT ROI DCNN 1 Krizhevsky ROI DCNN SVM ROI ROI DCNN DCNN ROI 8 [pixel] ROI SVM CT ROI ROI CT 4. 2 DCNN LOPOCV 3

4 1 CT ROI DCNN SVM Consolidation(CON) GGO Honeycomb (HCM) Reticular (RET) Emphysema (EMP) Nodular (NOD) Normal (NOR) pool Pool SVM DCNN Margin Normal Vector Support Vector Separatimg Hyperplane Train Data Separatimg Hyperplane 4.1 SVM DCNN SVM 4 SVM 1 SVM Train Data Test Data 4 Train Data SVM Test Data conv1, pool1, conv2, pool2, conv3, conv4, conv5, pool5, fc6 SVM 4 5 NOD GGO CT DCNN DCNN DCNN pool pool CT Pool Projection Test Data 4 SVM Train Data SVM Test Data SVM DCNN CT DCNN DCNN pool DCNN DCNN CT DCNN DCNN CT DCNN Caltech-101 SVM 6 DCNN 4

5 2 SVM LOPO DCNN conv1 pool1 conv2 pool2 conv3 conv4 conv5 pool5 fc6 Proposed 50.08% 70.07% 69.46% 70.24% 70.84% 72.67% 72.64% 75.55% 80.04% CT 50.96% 66.04% 67.57% 74.38% 74.61% 75.25% 77.55% 77.85% 74.13% Nat 52.27% 69.41% 68.15% 70.40% 66.79% 66.33% 69.38% 69.94% 75.78% Conv1 (Proposed) Pool1 (Proposed) Conv2 (Proposed) Pool2 (Proposed) Conv5 (Proposed) Pool5 (Proposed) Fc6 (Proposed) distance Conv1 (CT Img.) Pool1 (CT Img.) Conv2 (CT Img.) Pool2 (CT Img.) Conv5 (CT Img.) Pool5 (CT Img.) Fc6 (CT Img.) Conv1 (Nat.Img.) Pool1 (Nat.Img.) Conv2 (Nat.Img.) Pool2 (Nat.Img.) Conv5 (Nat.Img.) Pool5 (Nat.Img.) Fc6 (Nat.Img.) 5 Nodular GGO 2 SVM Nodular GGO 1-1 conv1 pool1 conv2 pool2 conv5 pool5 fc6 CT 1 3 DCNN 4.2 cos 4.1 DCNN 4.1 SVM (3) cos [10] cos p f p,i q f p,i s = f p,i f q,i / f p,i f q,i DCNN cos 4 4 DCNN 4 pool5 cos 3 SVM 5. DCNN DCNN DCNN

6 SVM DCNN conv1 pool1 conv2 pool2 conv3 conv4 conv5 pool5 fc6 Proposed CT Nat DCNN cos pool5 cos NOD EMP HCM CON GGO NOR RET conv pool conv pool conv conv conv pool fc Conv1 (Nat.Img.) Pool1 (Nat.Img.) Conv2 (Nat.Img.) Pool2 (Nat.Img.) Conv3 (Nat.Img.) Conv4 (Nat.Img.) Conv5 (Nat.Img.) Pool5 (Nat.Img.) Fc6 (Nat.Img.) 6 Caltech-101 Motorbikes Airplanes pool5 cos pool5 [1] K. Fukushima, Neocognitron: A Self-organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position, Biological Cybernetics, Vol. 36, No. 4, pp , [2] H. Shouno, Recent Studies around the Neocognitron, in Proc. Lecture Notes in Computer Science, Springer, Vol. 4985, pp , [3] Quoc V. Le, Marc Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg S. Corrado, Jeff Dean, Andrew Y. Ng, Building High-level Features Using Large Scale Unsupervised Learning, Acoustics, Speech and Signal Processing, pp , [4],,, Deep Convolutional Neural Network,, vol. 114, no. 515, NC , pp , [5] Li Deng, Dong Yu, Deep Learning: Methods and Applications, Microsoft Research, [6] Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton, ImageNet Classification with Deep Convolutional Neural Networks, Advances in Neural Information Processing Systems 25, [7] Dominik Scherer, Andreas Müller, and Sven Behnke, Evaluation of Pooling Operations in Convolutional Architectures for Object Recognition, Artificial Neural Networks ICANN 2010 Lecture Notes in Computer Science Volume 6354, 2010, pp [8] C. M. Bishop, :,,,,, Pattern recognition and machine learning-,, [9] Chih-Chung Chang, Chih-Jen Lin, LIBSVM: a library for support vector machines, ACM Transactions on Intelligent Systems and Technology (TIST), Vol. 2, No. 3, pp. 27, [10],,,,,,, 14(4), pp ,

Convolutional Neural Network A Graduation Thesis of College of Engineering, Chubu University Investigation of feature extraction by Convolution

Convolutional Neural Network A Graduation Thesis of College of Engineering, Chubu University Investigation of feature extraction by Convolution Convolutional Neural Network 2014 3 A Graduation Thesis of College of Engineering, Chubu University Investigation of feature extraction by Convolutional Neural Network Fukui Hiroshi 1940 1980 [1] 90 3