307 (2017.2.27) 307-8 Deep Convolutional Neural Network X Detecting Masses in Mammograms Based on Transfer Learning of A Deep Convolutional Neural Network Shintaro Suzuki, Xiaoyong Zhang, Noriyasu Homma, Makoto Yoshizawa Tohoku University : (Neural networks) (Image processing) (Deep learning) (Transfer learning), (Mammogram), (Computer-aided diagnosis/detection) : 980-8579 6-6-05 ( ) Tel.: (022)795-7130 E-mail: shintaro.suzuki.r4@dc.tohoku.ac.jp 1. 1 1) X X X 2 computer-aided diagnosis: CAD 2) X Fig. 1 X CAD deep learning deep convolutional neural network (DCNN) 1
Fig. 1: X DCNN X DCNN X DCNN DCNN DCNN DCNN X 2. 2.1 DCNN 3 DCNN 3 R,G,B DCNN fc8 ImageNet 4) 1000 Fig. 2 conv1-5 pool1,2,5 fc6-8 AlexNet 3) 2.2 DCNN 5) DCNN X Fig. 3(a) 120 ImageNet 4) DCNN DCNN 2 Fig. 3(b) X ROI fine-tuning DCNN Krizhevski AlexNet 3) Fig. 2 DCNN 5 2
Fig. 2: DCNN 3. 3.1 (fine-tuning) X Digital Database for Screening Mammography (DDSM) 6) region of interest (ROI) 454 227 ROI 885 969 1, 854 9 3.2 (fine-tuning) Fig. 4 50 6 % 10 % 50 DCNN Table 1 92.5 % 12.3 % Fig. 5 MASNOR Fig. 5(a) Fig. 5(b) Fig. 6 Fig. 6(a) Fig. 6(b) ROI ROI Fig. 7 receiver operating 3
(Mass) (Normal) Natural images (a) Mammograms (b) X fine-tuning Fig. 3: DCNN 30 25 訓練誤差 テスト誤差 Table 1: 50 20 15 10 5 819 119 938 66 850 916 885 969 1854 0 0 10 20 30 40 50 Fig. 4: fine-tuning characteristic (ROC) area under the curve: AUC 1 AUC 0.97 Kom et al. 7) AUC = 0.93 Sahier et al. 8) AUC = 0.87 90% 9.0% 24.5% 95% 18.7% 48.0% 4. X DCNN DCNN DCNN DCNN X DCNN ROI 1) A. Matsuda et al., Cancer Incidence and Incidence Rates in Japan in 2008: A Study of 25 Population-based Cancer Registries for the Monitoring of Cancer Incidence in Japan (MCIJ) Project, 4
(a) (b) Fig. 5: MAS: NOR: (a) (b) 5
真陽性率 [%] (a) (b) Fig. 6: (a), (b) 100 90 80 70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 偽陽性率 [%] 転移学習 ( 提案法 ) 医用画像のみ 70 80 90 Fig. 7: ROC 100 Japanese Journal of Clinical Oncology, 44-4, 388/396 (2013) 2) J. Tang et al, Computer-Aided Detection and Diagnosis of Breast Cancer With Mammography: Recent Advances, IEEE Trans. Inf. Technol. Biomed, 13-2, 236/251 (2009) 3) A. Krizhevski et al., ImageNet Classification with Deep Convolutional Neural Networks, In Proc. NIPS (2012) 4) J. Deng et al., ImageNet: A Large- Scale Hierarchical Image Database, In CVPR09, 248/255 (2009) 5) S. J. Pan and Q. Yang: A Survey on Transfer Learning, Knowledge and Data Engineering, IEEE Trabsaction, 22-10, 1345/1359 (2010) 6) M. Heath et al, The Digital Database for Screening Mammography, In Proceedings of the Fifth International Workshop on Digital Mammography, M.J. Yaffe, ed., pp. 212-218, Medical Physics Publishing, 2001. 7) G. Kom, A et al., Automated detection of masses in mammograms by local adaptive thresholding, Comput. Biol. Med., 37-1, 37/48 (2007) 8) B. Sahiner et al, Classification of mass 6
and normal breast tissue: A convolutional neural network classifier with spatial domain and texture images, IEEE Trans. Med. Imag., 15-5, 598/610 (1996) 9) N. R. Mudigonda et al, Gradient and texture analysis for the classification of mammographic masses, IEEE Trans. Med. Imag., 19-10, 1032/1043 (2000) 7