의료영상에서의딥러닝 - ksiim.org

1 (machine learning) . , . , , . (deep learning) . (artificial neural network) , . 1970 , . 2014;20:13-18-13- Deep Learning in Medical ImagingHansang Lee, Minseok Park, Junmo KimDepartment of Electrical Engineering, KAIST, Daejeon, Korea= Abstract =As a branch of artificial intelligence, machine learning has been one of the most important technolo-gies not only in computer vision but also in medical image analysis.

2 Currently, the deep learning, aspecific model of artificial neural networks, is getting increasing attention with its theoretical noveltyand significantly successful performance in computer vision. Furthermore, since deep learning hasshifted the paradigm of feature extraction from the hand-crafted features to the learnt-from-data fea-tures, deep learning is considered as a promising machine learning framework in medical imaging ap-plications. In this paper, we briefly introduce the concepts of deep learning and its several , we review the related works on deep learning applications in medical image analysis in-cluding image classification, detection and words: Machine learning, Deep learning, Artificial neural network, Medical imaging, Image analysis : , (305-338) 291 Tel: 042-350-3488, Fax: 042-350-3410E-mail.

3 2 . 3 . 4 . 1.. 1943 McCulloch and Pitts [1] . 1949 Hebb [2] , Hebbian learning . , , Hebbian learning . , .2. 1958 Rosenblatt [3, 4] (per-ceptron) 1 . (input layer) (output layer) (feed-forward network) . (weight).

4 , (threshold) (activation func-tion) . - , XOR , 1969 Minsky and Papert [5] . (hid-den layer) (multi-layer percep-tron) , , Minsky and Papert [5] 1970 .3. 1980 (back-propagation algorithm) . , (gradient descent).

5 1969 1986 Rumelhart, Hinton and Williams[6] .. (global minimum) (local minimum) , , , (super-vised learning) (overfitting) . 1995 (support vector machine, SVM) . 2014;20:13-18-14-4. 3 2006 Hinton (restricted Boltzmannmachine, RBM) DBN(deep belief network)[7, 8] . , . (generativemodel) , RBM.

6 RBM (feature extraction) , (unsupervised learning) .DBN , RBM . DBN RBM (pre-training) , , (fine-tuning) .. S (sigmoid unit) (rectified linearunit, ReLU) [28] . (vanishing gradient problem) . 2013 Dahl (drop-out) [10] , (convolutional neural network, CNN) [11] . (GPU) (big data).

7 5.. , , DBN . (deterministic model) CNN . 1 .. , SIFT [12] SURF [13] , (hand-crafted feature) . , , . : -15- 1. - [7] DBN [8]Hinton (stochastic model)- - Rumelhart (deterministic model) [11]LeCun.

8 , , . SIFT . 1. (manifold learning) . Sx, Sy, Sz Sx Sy Sz . , (manifold) . (locally linear embedding, LLE) [14], (Laplacian eigenmaps, LEM) [15], (Isomaps) [16] , (proximity graph) . Gerber [17].

9 Brosch [18] DBN [8] . RBM DBN , RBM (convRBM) [19] . , , . 300 (Alzheimer s disease, AD) T1 ADNI [20] , 128 128 128 2 2 . 2 (ventricle) AD .2. (basal-cell carcinoma, BCC) . (histopathology image) , , , . (discrete cosine transform, DCT), (wavelet) (Gabor descriptor) , BOF(bag of features) [21, 22].

10 , . Cruz-Roa [23] . RBM (autoencoder)[24] , (convolution) (pooling) (softmax) [25] , . 1417 BOF . , (digital staining) 2014;20:13-18-16- .3.. SIFT Haar . Kim [26].