29 CNN ( Extraction and Classification of Cell Nuclei Using CNN Features)

29 CNN ( Extraction and Classification of Cell Nuclei Using CNN Features) 28 28414063

i 1 1 2 3 3 5 3.1............................. 5 3.2 Support Vector Machine(SVM)............................. 6 3.2.1 SVM..................................... 6 3.2.2 SVM.................................... 9 3.3 Histograms of Oriented Gradients............................ 9 3.3.1.................................. 10 3.3.2............................ 10 3.3.3.............................. 10 3.4 Convolutional Neural Network............................. 11 3.4.1.................................... 11 3.4.2................................... 13 3.5 AlexNet............................... 13 4 15 4.1....................................... 15 4.1.1............................... 15 4.1.2......................... 15 4.2................................. 17 4.3........................... 19 4.3.1.................................. 19 4.3.2................................... 20 4.3.3.................................... 21 5 23 5.1.......................................... 23 5.2.......................................... 24 5.3.......................................... 24 6 26 6.1....................................... 26 6.1.1............................... 26

ii 6.1.2......................... 27 6.2................................ 28 6.2.1............................... 28 7 30 7.1.......................................... 30 7.2.......................................... 30 7.2.1............................ 31 7.2.2....................... 33 7.3............................... 36 8 38 39 40 42

1 1 [1]. 2405 [2] 5 1 1 [3][4] [5] [5] Higher-order Local AutoCorrelation HLAC

1 2 [6], Deep Learning Convolutional Neural Network(CNN). Hematoxylin-Eosin HE CNN

3 2 [7][8][9] [7][8] 3 2 G [9] [10][11] [10] Genetic Programing GP [12] Simulated Annealing SA [13] Simulated Annealing Programming SAP GP-SAP [10]

2 4 [11] Support Vector Machine(SVM) [11] HE HE [11] SVM HE

5 3 3.1 (HE) 3.1 3.1: HE

3 6 3.2 Support Vector Machine(SVM) SVM, 2 SVM. 2. 3.2.1 SVM C 1, C 2 2. 3.2. 3.2: 2 3.3 ( ) ( )

3 7 3.3 3.3: SVM,., x 1,, x n, y 1,, y n x i C 1 y i = 1, x i C 2 y i = 1 SVM f(x i ) = sign(g(x i )) g(x i ) = w x i + b (3.1) w b g(x i ) = 0 D(x i ) 0 C 1 g(x i ) < 0 C 2 x +s x s w b

3 8 2 (w x +s ) + b = +1 (w x s ) + b = 1 2 w (3.2) x +s x s x i y i (w t x i + b) 1 0 y i (w t x i + b) 1 0 (3.3) L(w) = 1 2 w 2 (3.4) w λ n λ i y i = 0, λ i 0 (i = 1,, n) (3.5) i=1 F (λ i ) = n λ i 1 n λ i λ j y i y j x i x j (3.6) 2 i=1 i,j=1 λ λ g(x i ) = n λ i y i x i x + b (3.7) i s b = y s n λ i y i x i x s (3.8) i s

3 9 3.2.2 SVM XOR ϕ x i x j ϕ(x i ) ϕ(x j ) SVM SVM 3.3 Histograms of Oriented Gradients Histograms of Oriented Gradients(HOG)[14] HOG HOG 3.4 3 3.4: HOG

3 10 3.3.1 m θ m(x, y) = f x (x, y) 2 + f y (x, y) 2 (3.9) θ = tan 1 f y(x, y) f x (x, y) (3.10) f x (x, y) = I(x + 1, y) I(x 1, y) f y (x, y) = I(x, y + 1) I(x, y 1) (3.11) I(x, y) (x, y) 3.3.2 m θ 0 180 20 9 3.3.3 1 1 60 60 1 5 5 1 3 3 10 10 100 HOG 100 9 (3 3) =8100

3 11 3.4 Convolutional Neural Network Convolutional Neural Network(CNN) Deep Learning CNN 3.5 CNN CNN 3.5: CNN 3.4.1 CNN feed-forward (3.12) a (k) ij = m 1 s=0 n 1 t=0 w (k) s t x (i+s)(j+t) + b (k) (3.12) m n ( ) x k a (k) 2 w (k) b (k) E M, N backpropagation

3 12 (3.13) E w (k) s t = M m i=0 N n j=0 E a (k) ij a (k) ij w (k) s t = M m i=0 N n j=0 E a (k) ij x (i+s)(j+t) M m E b = (k) i=0 N n j=0 E a (k) ij a (k) ij b (k) = M m i=0 N n j=0 E a (k) ij (3.13) backpropagation (3.14) δ (k) ij := E a (k) ij (3.14) Rectified Linear Unit( ReLU) ReLU feed-forward ReLU (3.15) a ij = ReLU (x ij ) = max (0, x ij ) (3.15)

3 13 backpropagation (3.16) E x ij = E a ij if a ij 0 0 otherwise (3.16) ReLU 3.4.2 max pooling mean pooling max pooling feed-forward backpropagation feed-forward (3.17) backpropagation (3.18) a ij = max (x (li+s)(lj+t) ) where s [0, l], t [0, l] (3.17) E x (li+s)(lj+t) = E a ij if a ij = x (li+s)(lj+t) 0 otherwise (3.18) 3.5 AlexNet CNN

3 14 AlexNet [15] CNN ILSVR2012 3.6 3.6: AlexNet

15 4, SVM 4.1 SVM 4.1.1 HE 6.1 600 ( 300 300 ) 400 ( 200 200 ) 1000 4.1.2 HOG CNN

4 16 (a) (b) 4.1: 3 CNN Alex-net 7 4096 3 7 3 10 5 7.1 4.1: HOG& 0.892 0.886 0.902 CNN 0.967 0.941 0.954 0.972 0.952 0.959 7.1, CNN HOG CNN

4 17 CNN CNN 1 4.2 HE SVM 4.2 CNN CNN [11] CNN SVM 90px 90px, 5px CNN SVM 2 SVM 4.3

4 18 4.2: (a) 4.3: (b)

4 19 4.3 4.3b Step 1: Step 2:. Step 3:. 4.3.1 4.1 4.4 P P = 2 (S red(i)) (4.1) S red(i) I RGB R RGB R G B 3 4 R RGB R

4 20 4.4: 4.3.2 Step 1: 2. Step 2: 2. Step 3:. Step 4: 2 Step 5: Step 4. 4.5

4 21 (a) 4.5: (b) 4.3.3 HE Step 4 Step 1: 4.3.1 Step 2: Step 1 2. Step 3:. Step 4: Step 5: 2 ( )

4 22 500 500 500

23 5. 5.1 HE 5.1 1920 1440 (a) (b) 5.1:

5 24 5.2 5.2 5.3 5.2a 5.3a (a) 5.2: (b) (a) 5.3: (b) 5.3 Precision Recall F-measure TP

5 25 FP FN True Positive False Positive False Negative T P 80, F P F N. T P P recision = T P + F P T P Recall = T P + F N 2 P recision Recall F measure = P recision + Recall (5.1) (5.2) (5.3) 7.1 5.1: Precision Recall F-measure 0.7466 0.8794 0.8076 0.6878 0.9198 0.7673 7.1 F 8 Recall

26 6 6.1 SVM 6.1.1 DNA DNA 2 1

6 27 6.1 800 800 1600 (a) (b) 6.1: 6.1.2 CNN 2 800 800 7 3 10 5

6 28 7.1 6.1: 0.723 CNN 0.891 7.1 CNN CNN 6.2 CNN Step 1: Step 2: Step 3: CNN Step 4: SVM Step 5: Step 2 Step 4 6.2.1 CNN 6.2a 6.2b CNN Step 1: x y

6 29 (a) CNN (b) CNN 6.2: CNN Step 2:, a b, Step 3: a 90 6.3a 90 90 Step 3 : a>90 6.3b a a 90 90 (a) a 90 (b) a>90 6.3:

30 7. 7.1 HE 3 HE 1920 1440 7.2 1 1 16 2

7 31 7.2.1 20 40 20 80 7.1 7.2 7.3 7.1 (a) (b) 7.1: (a) (b) 7.2:

7 32 (a) (b) 7.3:

7 33 7.1: 20 0 1.0 15 25 0.63 0.82 0 20 1.0 7.1 0.82 2 2 7.2.2 2 2 2 6 7.4 7.5 7.6 3 N/C N/C N/C 1 7.2

7 34 7.4: 7.5:

7 35 7.6:

7 36 7.2: NC 1 1780 581 0.23 2 1869 430 0.22 1 1089 866 0.28 2 813 874 0.28 1 499 1814 0.31 2 421 1905 0.30 7.2 N/C 7.3 N/C 2 CNN 2 2 ( ) 0 IV 5 IV

7 37 ( ) [16][17]

38 8 HE CNN SVM HE CNN,

39

40 [1] Nakhleh, R., Coffin, C., Cooper, K: Recommendations for quality assurance and improvement in surgical and autopsy pathology Hum Pathol, 37, pp. 985-988, 2006. [2] 2017 10 1 [3] Taylor, C.R., Levenson, R.M: Quantification of immunohistochemistry issues concerning methods, utility and semiquantitative assessment II Histopathology, 49, pp. 411-424, 2006. [4] : The IEICE transactions on information and systems (Japanese edetion) 96(4), pp. 782-790, 2013 [5], : MPS 2010-MPS-81(32), pp. 1-6, 2010 [6] : Proceedings of the Japan Joint Automatic Control Conference 57(0), pp. 1105-1107, 2014 [7],,,,. D-II,, II-. J77-D-2(2), pp. 449-452, 1994. [8],,,: Medical Imaging Technology 14(1), pp. 23-30, 1996.

41 [9],,,,,. MBE, ME. 112(123), pp. 31-34, 2012. [10] : GP SAP MPS 2009-MPS-75(12), pp. 1-6, 2009 [11],,,,, KONICA MINOLTA TECHNOLOGY REPORT. Vol. 13, 2016. [12] John R. Koza, Genetic programming, on the programming of computers by means of natural selection MIT Press, 1992 [13] Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N. Rosenbluth, Augusta H. Teller, Edward Teller, Equation of state calculation by fast computing machines The Journal of Chemical Physics 21, pp.1087, 1953 [14] Navneet Dalal, Bill Triggs, Histograms of oriented gradients for human detection, Proc. of IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pp. 886-893, 2005. [15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks, In Advances in neural information processing systems, pp. 1097-1105, 2012. [16] Marc Amoyel, Erika A. Bach. Cell competition: how to eliminate your neighbours, Development2014, Development 2014 141: pp, 988-1000, 2014 [17] Hogan C., Kajita M., Lawrenson K., Fujita Y. Interactions between normal and transformed epithelial cells: their contributions to tumourigenesis, Int J Biochem Cell Biol 43(4), pp.496-503, 2011

42 1.,,,,, CNN, 42, 2017. 2. Yuya Tsukada, Yuji Iwahori, Kenji Funahashi, Mami Jose, Jun Ueda, Takashi Iwamoto, Extraction of Cell Nuclei using CNN Features, Procedia Computer Science, Elsevier, Vol.112, Pages 1633-1640, 2017.