Presentation in Kyoto University Hospital

Presentation in Kyoto University Hospital

Transcript

1. ਓ޻஌ೳߨԋձ2019 େࡕࢢཱେֶେֶӃҩֶܥݚڀՊɹ์ࣹઢ਍அֶɾIVRֶڭࣨ ਓ޻஌ೳݚڀࣨɹ২ాେथ ਓ޻஌ೳͷ֓ཁ͔Β ҩྍԠ༻·Ͱ ※ݸਓར༻ͷൣғͰ͓ئ͍ਃ্͛͠·͢ɻ
2. ϝσΟΞܝࡌྺ ɾ೴ಈ຺ᚅ ே೔৽ฉ ೴݂؅ͷʮ͜ͿʯɺAIͰݕग़཰޲্ େࡕࢢେͳͲ
 https://www.asahi.com/articles/ASLBT3T0VLBTUBQU004.html CB newsɺYahoo!χϡʔε ೴ಈ຺ᚅɺAIิॿͰ์ࣹઢՊҩͷಡӨਫ਼౓޲্ ਆށ৽ฉ
 https://headlines.yahoo.co.jp/hl?a=20181025-14400000-cbn-soci ɾϚϯϞάϥϑΟ 
 ே೔৽ฉ ೕ͕Μͷݟམͱ͠ɺਓ޻஌ೳͰ๷͙ େࡕࢢཱେ͕։ൃ https://www.asahi.com/articles/ ASL4C22S8L4CUBQU003.html த೔৽ฉ ͜ͷਓ ࢈ܦ৽ฉ େࡕࢢେ͕AI׆༻ͨ͠ೕ͕Μͷը૾਍அγεςϜ։ൃ https://www.sankei.com/west/news/180413/wst1804130090-n1.html
3. λΠϜϥΠϯ σΟʔϓϥʔχϯάͷҩྍԠ༻ σΟʔϓϥʔχϯά֓࿦ σΟʔϓϥʔχϯάͱҩࢣͷೝࣝͷࠩ σΟʔϓϥʔχϯάͱҩྍͷະདྷ σΟʔϓϥʔχϯάͷ࣭ͷอূ
4. ɾ͋Γͱ͋ΒΏΔಛ௃Λநग़Ͱ͖Δ ɾಛ௃ྔநग़ͷΩϞͷٕज़͸৞ΈࠐΈ૚Ͱ͋Δ ɾ൒ࣗಈͰಛ௃ྔநग़Λߦ͑Δ ·ͱΊ: σΟʔϓϥʔχϯά֓࿦
5. ਓ޻஌ೳ(AI)ͱ͸ େྔͷ஌ࣝσʔλʹରͯ͠ɺ ߴ౓ͳਪ࿦Λత֬ʹߦ͏͜ͱΛ໨ࢦͨ͠΋ͷ Ұൠࣾஂ๏ਓ ਓ޻஌ೳֶձΑΓൈਮ
6. ਂ૚ֶश(σΟʔϓϥʔχϯά)ͱ͸ ଟ૚ͷχϡʔϥϧωοτϫʔΫΛ༻ֶ͍ͨश
7. ࣮ࡍ͸ ʮਓ޻஌ೳʯ =ʮσΟʔϓϥʔχϯά(ਂ૚ֶश)ʯ ͱ༻͍ΒΕ͍ͯΔ͜ͱ͕ଟ͍ɻ
8. χϡʔϥϧωοτϫʔΫͱ͸ ͦͷχϡʔϩϯ΁ લͷχϡʔϩϯ͔Β৴߸͕ೖྗ͞ΕΔ Ұఆͷᮢ஋Λ௒͑Δͱ ͦͷχϡʔϩϯ͸ൃՐ࣍͠΁৴߸ΛૹΔ શ͔ແ͔ͷ๏ଇ
9. χϡʔϥϧωοτϫʔΫͱ͸ ͦͷχϡʔϩϯ΁ લͷχϡʔϩϯ͔Β৴߸͕ೖྗ͞ΕΔ Ұఆͷᮢ஋Λ௒͑Δͱ ͦͷχϡʔϩϯ͸ൃՐ࣍͠΁৴߸ΛૹΔ શ͔ແ͔ͷ๏ଇ
10. χϡʔϥϧωοτϫʔΫͱ͸ x x y 1 2 1 2 w w y = x w + x w 2 2 1 1 શ͔ແ͔ͷ๏ଇ ͦͷχϡʔϩϯ΁ લͷχϡʔϩϯ͔Β৴߸͕ೖྗ͞ΕΔ Ұఆͷᮢ஋Λ௒͑Δͱ ͦͷχϡʔϩϯ͸ൃՐ࣍͠΁৴߸ΛૹΔ h(x) = 0 (y ≦ 0) 1 (y > 0) {
11. χϡʔϥϧωοτϫʔΫͱ͸ ̍࣍ؔ਺ͱͯ͠(࠷খ୯Ґͷ)ͷχϡʔϥϧωοτϫʔΫ
12. χϡʔϥϧωοτϫʔΫͱ͸ →௚ઢతʹ෼ྨͰ͖Δ؆୯ͳಛ௃நग़ʹ༗ޮ ಛ௃நग़Մೳ ಛ௃நग़ෆՄೳ ̍࣍ؔ਺ͱͯ͠(࠷খ୯Ґͷ)ͷχϡʔϥϧωοτϫʔΫ
13. σΟʔϓϥʔχϯάͱ͸ ଟ࣍ݩؔ਺Ͱද͞ΕΔ
14. σΟʔϓϥʔχϯάͱ͸ →ཧ࿦্ɺ“͋Γͱ͋ΒΏΔಛ௃நग़” ʹ༗ޮ ଟ࣍ؔ਺ͱͯ͠ͷσΟʔϓϥʔχϯά ಛ௃நग़Մೳ ಛ௃நग़Մೳ
15. ·ͱΊ: σΟʔϓϥʔχϯά֓࿦ ɾ͋Γͱ͋ΒΏΔಛ௃Λநग़Ͱ͖Δ ɾಛ௃ྔநग़ͷΩϞͷٕज़͸৞ΈࠐΈ૚Ͱ͋Δ ɾ൒ࣗಈͰಛ௃ྔநग़Λߦ͑Δ
16. ࣮ࡍʹͲ͏΍ͬͯಛ௃ྔநग़͍ͯ͠Δͷ͔ʁ
17. σΟʔϓϥʔχϯάͷਪ࿦ Input Output
18. Waldrop MM. News Feature: What are the limits of deep learning? Proceedings of the National Academy of Sciences. 2019;116(4):1074-7. σΟʔϓϥʔχϯάͷਪ࿦ ֤૚Λ ՄࢹԽ
19. Waldrop MM. News Feature: What are the limits of deep learning? Proceedings of the National Academy of Sciences. 2019;116(4):1074-7. σΟʔϓϥʔχϯάͷਪ࿦ ઙ͍૚ ਂ͍૚ ୯७ͳܗ ෳࡶͳܗ Ұ࣍ؔ਺ ଟ࣍ؔ਺ →֤૚͸ϑΟϧλʹͳ͍ͬͯΔ →ਓؒͷࢹ֮໺Λ࠶ݱͨ͠
20. ਤ ϑΟϧλ
21. σΟʔϓϥʔχϯάͷਪ࿦ ద੾ͳϑΟϧλΛ࡞੒͍ͯ͘͠࡞ۀ͕ͦ͜ ಛ௃ྔͷநग़Ͱ͋Δɻ
22. ࣮ࡍʹͲ͏΍ͬͯϑΟϧλ࡞੒͍ͯ͠Δͷ͔ʁ →໢ཏతݕࡧ(ྗٕ)Ͱ͢ɻ
23. σΟʔϓϥʔχϯάͷֶश Training dataset Validation dataset Test dataset Developing dataset ֶशʹ༻͍Δ ↓
24. Training༻ Validation༻ ֶश ϑΟʔυόοΫ σΟʔϓϥʔχϯάͷֶश
25. σΟʔϓϥʔχϯάͷ࢓૊Έ Input layer Output layer Convolutional layer Convolutional layer Convolutional layer Convolutional layer = ৞ΈࠐΈ૚ = ϑΟϧλʔ
26. Convolutional layer 1 1 1 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 0 0 1 1 1 1 0 ֶश͍ͨ͠ಛ௃
27. Convolutional layer 1 1 1 1 1 1 1 1 1 3x3 ͰॏΈ͕ͦΕͧΕ1ͷ৞ΈࠐΈ
28. Convolutional layer 1 1 1 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 0 0 1 1 1 1 0
29. Convolutional layer 1 1 1 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 0 0 1 1 1 1 0 6
30. Convolutional layer 1 1 1 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 0 0 1 1 1 1 0 6 6
31. Convolutional layer 1 1 1 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 0 0 1 1 1 1 0 6 6 3 5 5 2 6 5 2
32. Convolutional layer
33. Convolutional layer 0 1 1 1 1 1 1 1 1 3x3 ͰॏΈ͕ࠨ্͚ͩ0ͷ৞ΈࠐΈ
34. Convolutional layer
35. Convolutional layer 1 0 1 1 1 1 1 1 1 3x3 ͰॏΈ͕ਅ্͚ͩ0ͷ৞ΈࠐΈ
36. Convolutional layer 1 1 1 0 0 0 1 1 1 0 0 1 0 0 0 1 1 0 0 0 1 1 1 1 0
37. Deep learningͷֶश 1 0 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
38. Deep learningͷֶश ϑΟϧλ Training data Validation data ʹ ͜ͷࠩҟΛϑΟϧλͷඍௐ੔ͰͰ͖Δ͚ͩখ͘͞ ͍ͯ͘͜͠ͱ͕Deep learningͷֶशաఔ ෼ྨ-Valid ෼ྨ-Train
39. ແ਺ύλʔϯͷϑΟϧλʔ(৞ΈࠐΈ)Λ࡞Γɺ ϑΟϧλʔ(৞ΈࠐΈ)ΛมԽͤͨ͞ͱ͖ͷਖ਼ղ͔Βͷಈ͖Λ ඍ෼͠ɺ࠷దղΛಋ͍͍ͯ͘
40. ɾ͋Γͱ͋ΒΏΔಛ௃Λநग़Ͱ͖Δ ɾಛ௃ྔநग़ͷΩϞͷٕज़͸৞ΈࠐΈ૚Ͱ͋Δ ɾ൒ࣗಈͰಛ௃ྔநग़Λߦ͑Δ ·ͱΊ: σΟʔϓϥʔχϯά֓࿦
41. from keras.applications.inception_v3 import InceptionV3 from keras.preprocessing import image from keras.models import Model from keras.layers import Dense, GlobalAveragePooling2D from keras.optimizers import SGD from keras import backend as K base_model = InceptionV3(weights='imagenet', include_top=False) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) predictions = Dense(200, activation='softmax')(x) model = Model(inputs=base_model.input, outputs=predictions) for layer in base_model.layers: layer.trainable = True model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy') model.fit_generator(...) શίʔυ঺հ
42. શίʔυ঺հ ͨͬͨे਺ߦͷίʔυͰɺ ύϫϑϧͳσΟʔϓϥʔχϯά͕࢖͑Δɻ →ϥΠϒϥϦ TensorFlow(Google), Kaffe(Microsoft), Chainer(PreferredNetworks) →ϥούʔϥΠϒϥϦ Keras
43. from keras.applications.inception_v3 import InceptionV3 from keras.preprocessing import image from keras.models import Model from keras.layers import Dense, GlobalAveragePooling2D from keras.optimizers import SGD from keras import backend as K base_model = InceptionV3(weights='imagenet', include_top=False) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) predictions = Dense(200, activation='softmax')(x) model = Model(inputs=base_model.input, outputs=predictions) for layer in base_model.layers: layer.trainable = True model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy') model.fit_generator(...) શίʔυ঺հ ͕͜͜ɺϥΠϒϥϦͷҖྗʂ
44. from keras.applications.inception_v3 import InceptionV3 from keras.preprocessing import image from keras.models import Model from keras.layers import Dense, GlobalAveragePooling2D from keras.optimizers import SGD from keras import backend as K base_model = InceptionV3(weights='imagenet', include_top=False) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) predictions = Dense(200, activation='softmax')(x) model = Model(inputs=base_model.input, outputs=predictions) for layer in base_model.layers: layer.trainable = True model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy') model.fit_generator(...) શίʔυ঺հ ΞϧΰϦζϜΛ૊Ή
45. from keras.applications.inception_v3 import InceptionV3 from keras.preprocessing import image from keras.models import Model from keras.layers import Dense, GlobalAveragePooling2D from keras.optimizers import SGD from keras import backend as K base_model = InceptionV3(weights='imagenet', include_top=False) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) predictions = Dense(200, activation='softmax')(x) model = Model(inputs=base_model.input, outputs=predictions) for layer in base_model.layers: layer.trainable = True model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy') model.fit_generator(...) શίʔυ঺հ ΞϧΰϦζϜΛ࠷దԽ͢Δํ๏Λઃఆ͢Δ
46. from keras.applications.inception_v3 import InceptionV3 from keras.preprocessing import image from keras.models import Model from keras.layers import Dense, GlobalAveragePooling2D from keras.optimizers import SGD from keras import backend as K base_model = InceptionV3(weights='imagenet', include_top=False) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) predictions = Dense(200, activation='softmax')(x) model = Model(inputs=base_model.input, outputs=predictions) for layer in base_model.layers: layer.trainable = True model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy') model.fit_generator(...) શίʔυ঺հ ࣮ࡍʹֶशελʔτ
47. ɾ͋Γͱ͋ΒΏΔಛ௃Λநग़Ͱ͖Δ ɾಛ௃ྔநग़ͷΩϞͷٕज़͸৞ΈࠐΈ૚Ͱ͋Δ ɾ൒ࣗಈͰಛ௃ྔநग़Λߦ͑Δ ·ͱΊ: σΟʔϓϥʔχϯά֓࿦
48. σΟʔϓϥʔχϯάݚڀͷऑ఺ ɾֶशσʔλΛ४උ͢Δ͕͔͔࣌ؒΔɻ ɹˠྔͱ࣭ͷߴ͍σʔλͰ͋Ε͹͋Δ΄ͲΑ͍ɻ ɹɹগ਺ͷσʔλʹ͸ऑ͍ɻ ɾ࣮ࡍʹ΍Δ·Ͱ݁Ռ͕ग़ͯ͘ΕΔ͔Θ͔Βͳ͍ɻ ɾϚγϯύϫʔ͕͍Δ ɹˠGPUΛੵΜͩϚγϯ͕ඞཁɻ ɾઐ໳஌͕ࣝඞཁ ɹˠઐ໳Ո(ΤϯδχΞ)͕͍Δ΄͏͕Α͍ɻ
49. λΠϜϥΠϯ σΟʔϓϥʔχϯάͷҩྍԠ༻ σΟʔϓϥʔχϯά֓࿦ σΟʔϓϥʔχϯάͱҩࢣͷೝࣝͷࠩ σΟʔϓϥʔχϯάͱҩྍͷະདྷ σΟʔϓϥʔχϯάͷ࣭ͷอূ
50. σΟʔϓϥʔχϯάͷ։ൃͰॏཁͳ͜ͱ Training dataset Validation dataset Test dataset ػցֶशք ΞϧΰϦζϜͷ τϨʔχϯά༻ ΞϧΰϦζϜͷ τϨʔχϯά༻ ΞϧΰϦζϜͷ
 ධՁ༻ ҩྍք ΞϧΰϦζϜͷ τϨʔχϯά༻ ΞϧΰϦζϜͷ
 ධՁ༻ ΞϧΰϦζϜͷ
 ධՁ༻ σʔλʹ͸3छྨ͋Δ →ݱঢ়ɺఆ͕ٛҟͳ͓ͬͯΓࠞಉ͞Ε͍ͯΔɻ →ҩྍքͰ͸࠷ۙinternal validation/external validationͱ࢖͍෼͚
51. λΠϜϥΠϯ σΟʔϓϥʔχϯάͷҩྍԠ༻ σΟʔϓϥʔχϯά֓࿦ σΟʔϓϥʔχϯάͱҩࢣͷೝࣝͷࠩ σΟʔϓϥʔχϯάͱҩྍͷະདྷ σΟʔϓϥʔχϯάͷ࣭ͷอূ
52. Classification Detection Segmentation Super Resolution Natural Language Processing σΟʔϓϥʔχϯάٕज़ͷछྨ GAN
53. σΟʔϓϥʔχϯάٕज़ͷछྨ Classification ը૾1ຕ͔Βྑѱੑͷಛ௃ྔΛநग़ͯ͠਍அ͢Δ ྫ͑͹
54. ౰ՊͰͷϚϯϞάϥϑΟը૾ͷClassification True positive rate False positive rate 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 AUC: 0.94, Sensitivity : 85.1%Specificity : 93.1% ѱੑ͔ɺͦ͏Ͱͳ͍͔ͷ2෼ྨ
55. ౰ՊͰͷϚϯϞάϥϑΟը૾ͷClassification Malignant Malignant
56. [1] Lévy D, Jain A. Breast mass classification from mammograms using deep convolutional neural networks. arXiv preprint arXiv:161200542. 2016. [2] Ribli D, Horváth A, Unger Z, Pollner P, Csabai I. Detecting and classifying lesions in mammograms with Deep Learning. Sci Rep. 2018;8(1):4165. [3] de Lima SM, da Silva-Filho AG, dos Santos WP. Detection and classification of masses in mammographic images in a multi-kernel approach. Comput Methods Programs Biomed. 2016;134:11-29. [4] Lotter W, Sorensen G, Cox D. A multi-scale cnn and curriculum learning strategy for mammogram classification. Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Springer, 2017; p. 169-77. [5] Jadoon MM, Zhang Q, Haq IU, Butt S, Jadoon A. Three-class mammogram classification based on descriptive CNN features. BioMed research international. 2017;2017. No Accuracy [%] AUC Comment [1] 0.93 Mass images only [2] 0.95, 0.85 [3] 0.94 Mass images only [4] 0.92 [5] 0.84 ౰Պ 0.94 0.94 ઌߦݚڀͱͷൺֱ
57. The algorithm with the deep learning techniques could accurately detect breast cancers in mammograms with an AUC of 0.94. After applying detection techniques (SSD-Keras, Yolo), clinical trial is preparing. Conclusion ※2018೥์ࣹઢՊֶձ૯ձൃදࢿྉ
58. σΟʔϓϥʔχϯάٕज़ͷछྨ Detection ը૾1ຕͷதͷजᚅͷॴݟ͔Βྑѱੑͷಛ௃ྔΛநग़ͯ͠਍அ͢Δ ྫ͑͹
59. ౰ՊͰͷϚϯϞάϥϑΟը૾ͷDetection
60. ײ౓ʹ͍ͭͯ Hospital test-dataset Clinic test-dataset ײ౓ 0.99 (224/225) 0.97 (37/38) എܠೕથೱ౓ผ ɹࢷ๱ 1.00 (35/35) 1.00 (2/2) ɹೕથ૊৫ࢄࡒ 1.00 (86/86) 0.95 (18/19) ɹෆۉҰߴೱ౓ 0.99 (73/74) 1.00 (12/12) ɹσϯεϒϨετ 1.00 (30/30) 1.00 (5/5) පมେ͖͞ผ ɹCarcinoma in situ 1.00 (17/17) 1.00 (4/4) ɹ1-10 mm 1.00 (37/37) 1.00 (6/6) ɹ11-20 mm 1.00 (80/80) 0.95 (19/20) ɹ21-50 mm 0.99 (82/83) 1.00 (8/8) ɹ>50 mm 1.00 (8/8) NA ॴݟผ ɹजᚅ 1.00 (103/103) 0.83 (5/6) ɹੴփԽ 1.00 (83/83) 1.00 (20/20) ɹہॴੑඇରশੑӄӨ 1.00 (74/74) 1.00 (11/11) ɹߏஙͷཚΕ 0.99 (92/93) 1.00 (6/6)
61. FROC Sensitivity 0.85 0.88 0.91 0.94 0.97 1.00 False positive per image 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10 Test dataset1 Test dataset2
62. σΟʔϓϥʔχϯάٕज़ͷछྨ Segmentation ը૾1ຕͷதͷϐΫηϧϨϕϧͰྑѱੑͷಛ௃ྔΛநग़ͯ͠਍அ͢Δ ྫ͑͹
63. ౰ՊͰͷڳ෦Ϩϯτήϯը૾ͷSegmentation
64. ౰ՊͰͷڳ෦Ϩϯτήϯը૾ͷSegmentation ※preliminaryͳstudyͰ͢ ์ࣹઢՊҩͷൃݟͰ͖Δ݁અͰ͸ײ౓90%௒͑ 5mmҎԼͷ݁અͰ͸ײ౓60%͘Β͍
65. σΟʔϓϥʔχϯάٕज़ͷछྨ Super Resolution ݪը૾Λѹॖ͠ɺѹॖը૾ͷڭࢣσʔλͱͯ͠ݪը૾Λ༻͍ɺ ͦͷѹॖͷಛ௃ྔΛநग़͢Δ ྫ͑͹
66. ౰ՊͰͷϚϯϞάϥϑΟੴփԽը૾ͷSuper Resolution
67. ౰ՊͰͷϚϯϞάϥϑΟੴփԽը૾ͷSuper Resolution
68. ౰ՊͰͷϚϯϞάϥϑΟੴփԽը૾ͷSuper Resolution
69. ౰ՊͰͷϚϯϞάϥϑΟੴփԽը૾ͷSuper Resolution CNR (Contrast to Noise Ratio), SCTF (System Contrast Transfer Function )
70. Deep LearningͷҩྍԠ༻࿦จ঺հ
71. ɾNo. of Data train: 32586ྫ test: 2452ྫ ɾPaper Training and Validating a Deep Convolutional Neural Network for Computer-Aided Detection and Classification of Abnormalities on Frontal Chest Radiographs ɾResult AUC(normal: 0.964, cardiomegaly: 0.875, effusion: 0.962, consolidation: 0.850, edema: 0.868, pneumothorax: 0.861) Classification ɾTrainingArchitecture GoogLeNet ɾData Abstract train: internal test: internal 2016/10/24
72. ɾPaper Deep Learning at Chest Radiography: Automated Classification of Pulmonary Tuberculosis by Using Convolutional Neural Networks ɾTrainingArchitecture AlexNet, GoogLeNet ɾResult AUC:0.99(Ensemble) AUC:0.98(Alex,GoogLeNet) ɾData Abstract train: multi test: multi ɾNo. of Data train: 857ྫ (Tb+:417) test: 150ྫ (Tb+:75) 2017/1/19 Classification
73. ɾPaper Performance of a Deep-Learning Neural Network Model in Assessing Skeletal Maturity on Pediatric Hand Radiographs ɾTrainingArchitecture ResNet 50 ɾResult mean difference: 0 years mean RMS: 0.63 years mean MAD: 0.50 years ɾData Abstract train: multi test: external & OpenData ɾNo. of Data train: 14036ྫ test: 1113ྫ 2017/8/24 Classification
74. ɾPaper Deep Learning with Convolutional Neural Network for Differentiation of Liver Masses ɾNo. of Data train: 1068ྫ test: 100ྫ ɾTrainingArchitecture Six convolutional layers ɾResult model triphasic Median accuracy of differential diagnosis 0.84 A-B/C-E AUC:0.92 ɾData Abstract train: internal test: internal 2017/8/14 Classification
75. ɾPaper Deep Learning for Staging Liver Fibrosis on CT: a pilot study ɾNo. of Data train: 396ྫ test: 100ྫ ɾTrainingArchitecture Four convolutional layers ɾResult spearman rank correlation coefficients: 0.48 AUC ≥ F2: 0.74, ≥ F3: 0.76, F4: 0.73 ɾData Abstract train: internal test: internal 2018/4/18 Classification
76. ɾPaper Liver Fibrosis: Deep Convolutional Neural Network for Staging by Using Gadoxetic Acid– Enhanced Hepatobiliary Phase MR Images ɾNo. of Data train: 534ྫ test: 100ྫ ɾTrainingArchitecture Three convolutional layers ɾResult spearman rank correlation coefficients: 0.63 AUC:0.84(F4), 0.84(F3), 0.85(F2) ɾData Abstract train: internal test: internal 2017/10/24 Classification
77. ɾPaper Hybrid 3D/2D Convolutional Neural Network for Hemorrhage Evaluation on Head CT ɾNo. of Data train: 10159ྫ(+: 901) test: 682ྫ(+: 92) ɾTrainingArchitecture mask R-CNN ɾResult AUC: 0.981, Sensitivity: 0.951, Specificity: 0.973 Dice scores IPH: 0.931, epidural/subdural: 0.863, SAH: 0.772 ɾData Abstract train: internal test: internal 2018/6/6 Detection
78. ɾPaper Use of 2D U-Net Convolutional Neural Networks for Automated Cartilage and Meniscus Segmentation of Knee MR Imaging Data to Determine Relaxometry and Morphometry ɾNo. of Data train: 158(external) test: 16ྫ(external) ɾTrainingArchitecture U-Net ɾResult Dice coefficients DESS: meniscus 0.867(95%CI:0.859, 0.875), cartilage 0.833(95%CI:0.821, 0.845) ɾData Abstract (DESS) train: external test: external 2018/1/9 Segmentation
79. ɾPaper 3D Deep Learning Angiography (3D-DLA) from C-arm Conebeam CT ɾResult Dice similariy coefficients͸98.0% ɾNo. of Data train: 43 (4.3million labeled voxels: vasculature, bone, and soft tissue) test: 8 ɾTrainingArchitecture ResNet ɾData Abstract train: internal 2018/1/26 Super-Resolution
80. ɾPaper Denoising of 3D magnetic Resonance Images with Multi-channel Residual Learning of Convolutional Neural Network ɾNo. of Data train: 20ྫ(IXI‐Hammersmith) test: IXI‐Hammersmith(10), IXI‐Guys(10), Brainweb(?) ɾTrainingArchitecture Multi-channel DnCNN(s and g) ɾResult PSNR, SSIMͷ྆ऀͰMCDnCNNs͕ྑ͍ ɾData Abstract train: external test: external (same as above and not same as above) 2018/7/4 Super-Resolution
81. ɾResult Detection ਫ਼౓:100% Segmentation Dice similarity coefficient: 0.86 ± 0.09 ɾPaper Clinical Evaluation of a Multiparametric Deep Learning Model for Glioblastoma Segmentation Using Heterogeneous Magnetic Resonance Imaging Data From Clinical Routine ɾNo. of Data test: 64ྫ ɾTrainingArchitecture DeepMedic ɾData Abstract train: pretrained(glioma 249ྫ) test: internal 2018/4/11 Detection + Segmentation
82. ɾPaper Deep Learning Approach for Evaluating Knee MR Images: Achieving High Diagnostic Performance for Cartilage Lesion Detection ɾNo. of Data train segmentation: 175ྫ test segmentation: ͳ͠(three-fold) train classification: 16075patch(+:- = 1982: 14093) test classification: 1320patch (+:- = 660:660) ɾTrainingArchitecture SegNet + VGG16 ɾResult Segmentation(Dice coefficients) femur: 0.96 ± 0.02, tibia: 0.95 ± 0.03 femoral cartilage: 0.81 ± 0.04, tibial cartilage 0.82 ± 0.04 Classification AUC: 0.914-0.917(ೈࠎͷਖ਼ৗorҟৗ) ɾData Abstract train: internal (detection) 2018/6/1 ※classification͸segͷΜͰfinetune Segmentation + Classification
83. ɾPaper Development and Validation of a Deep Learning System for Staging Liver Fibrosis by Using Contrast Agent–enhanced CT Images in the Liver ɾNo. of Data(Segmentation) train: 50ྫ(five-fold) ɾTrainingArchitecture U-Net ɾResult Dice similarity index: 0.92 ɾData Abstract train: internal test: multiple 2018/7/17 Segmentation + Classification
84. λΠϜϥΠϯ σΟʔϓϥʔχϯάͷҩྍԠ༻ σΟʔϓϥʔχϯά֓࿦ σΟʔϓϥʔχϯάͱҩࢣͷೝࣝͷࠩ σΟʔϓϥʔχϯάͱҩྍͷະདྷ σΟʔϓϥʔχϯάͷ࣭ͷอূ
85. + = ύϯμ ςφΨβϧ Deep learning ਓؒͷೝ஌ͱσΟʔϓϥʔχϯάͷೝ஌ ਓؒ ύϯμ ύϯμ ← → ← →
86. ਓؒͷೝ஌ͱσΟʔϓϥʔχϯάͷೝ஌ Expectations Reality Human answer Deep learning answer What it is https://hackernoon.com/the-implications-of-adversarial-examples-deep-learning-bits-3-4086108287c7
87. Reality Radiologist answer Deep learning answer What it is https://hackernoon.com/the-implications-of-adversarial-examples-deep-learning-bits-3-4086108287c7 ์ࣹઢՊҩͷೝ஌ͱσΟʔϓϥʔχϯάͷೝ஌ ະ஌ɾਅ࣮΁ͷ௅ઓ ҩྍ҆શ΁ͷݒ೦
88. λΠϜϥΠϯ σΟʔϓϥʔχϯάͷҩྍԠ༻ σΟʔϓϥʔχϯά֓࿦ σΟʔϓϥʔχϯάͱҩࢣͷೝࣝͷࠩ σΟʔϓϥʔχϯάͱҩྍͷະདྷ σΟʔϓϥʔχϯάͷ࣭ͷอূ
89. ΞϧΰϦζϜͷධՁ ᶃ“Test dataset”ͰΈΔ ᶄ“Developing dataset”ͰΈΔ
90. ΞϧΰϦζϜͷධՁ Test datasetͷछྨ Split-sample test-dataset Geographic test-dataset Temporal test-dataset ͳ͠ Validation onlyͷstudyɻ overfitting͕ओମɻ 1ͭͷσʔληοτΛ8:1:1ʹ ෼͚Δ͜ͱ͕ଟ͍ɻ ௿ ߴ ৴ པ ౓ ྫ͑͹ɺಉҰࢪઃ͕ͩɺ ৽ͨʹ༻ҙ͞Εͨσʔληοτ ࢪઃͷҧ͏σʔληοτ
91. ઌߦݚڀʹ͍ͭͯ Citation Target Modality The deep learning model Category Develop/Evaluate The type of test dataset [12] Abnormality CR GoogLeNet Classification Develop and evaluate Split-sample test [91] Tb CR GoogLeNet Classification Develop and evaluate Split-sample test※ [93] Emergency HMH CT GoogLeNet Classification Develop and evaluate Split-sample test SAI [94] Bone age CR Unknown Classification Evaluate Geographic test [95] Liver mass CT Six convolutional layers Classification Develop and evaluate Split-sample test [96] Bone age CR ResNet Classification Develop and evaluate Geographic test [97] Liver fibrosis CT Four convolutional layers Classification Develop and evaluate Split-sample test [98] Liver fibrosis MR Three convolutional layers Classification Develop and evaluate Split-sample test [99] Image classification MR AlexNet, GoogLeNet Classification Develop and evaluate Split-sample test [100] Elbow effusion CR DenseNet Classification Develop and evaluate Split-sample test [101] Sinusitis CR ResNet Classification Develop and evaluation Geographic test [102] Mammo density CR ResNet Classification Develop and evaluation Temporal test [103] Aneurysms MR ResNet Classification Develop and evaluate Geographic test [104] Hemorrhage CT mask R-CNN Detection Develop and evaluate Temporal test Segmentation [105] Breast cancer study1 MG Unknown (ViDi Red) Segmentation (not clear) Develop and evaluate Geographical test study2 MG Temporal test [106] Knee DESS MR U-net Segmentation Develop and evaluate Split-sample test T1rho Split-sample test [108] Glioblastoma MR DeepMedic Detection Evaluate Geographic test Segmentation [109] Meningioma MR DeepMedic Detection Evaluate Geographic test Segmentation [111] Angiography CT ResNet Segmentation Develop and evaluate Split-sample test [112] Left ventricular MR U-net Segmentation Develop and evaluate Geographic test [123] Glioma MR DeepMedic Segmentation Evaluate Geographic test ResNet Classification Develop Cross-validation only [124] Cartilage MR SegNet Segmentation Develop and evaluate Cross-validation only VGG Classification Split-sample test [125] Liver fibrosis CT U-net Segmentation Develop and evaluate Cross-validation only GeNet or NEMO Classification Geographic test [127] Lung cancer CR ResNet Segmentation Develop and evaluate Cross-validation only Classification Geographic test [128] Neck anatomy CT Faster R-CNN Detection Develop and evaluate Split-sample test FCN Segmentation
92. Developing datasetͰΈΔσΟʔϓϥʔχϯάͷ܏޲ Sensitivity 0 0.2 0.4 0.6 0.8 1 The number of training data 0 400 800 1200 1600 2000 2400 2800 3200 3600 4000 Liver mass Liver fibrosis (CT) by Yasaka Liver fibrosis (MR) Liver fibrosis (CT) by Choi IPH EDH/SDH SAH Aneurysm
93. λΠϜϥΠϯ σΟʔϓϥʔχϯάͷҩྍԠ༻ σΟʔϓϥʔχϯά֓࿦ σΟʔϓϥʔχϯάͱҩࢣͷೝࣝͷࠩ σΟʔϓϥʔχϯάͱҩྍͷະདྷ σΟʔϓϥʔχϯάͷ࣭ͷอূ
94. ҩྍ΁ͷσΟʔϓϥʔχϯά࣌୅ͷ౸དྷ ARTICLE OPEN Scalable and accurate deep learning with electronic health records Alvin Rajkomar 1,2, Eyal Oren1, Kai Chen1, Andrew M. Dai1, Nissan Hajaj1, Michaela Hardt1, Peter J. Liu1, Xiaobing Liu1, Jake Marcus1, Mimi Sun1, Patrik Sundberg1, Hector Yee1, Kun Zhang1, Yi Zhang1, Gerardo Flores1, Gavin E. Duggan1, Jamie Irvine1, Quoc Le1, Kurt Litsch1, Alexander Mossin1, Justin Tansuwan1, De Wang1, James Wexler1, Jimbo Wilson1, Dana Ludwig2, Samuel L. Volchenboum3, Katherine Chou1, Michael Pearson1, Srinivasan Madabushi1, Nigam H. Shah4, Atul J. Butte2, Michael D. Howell1, Claire Cui1, Greg S. Corrado1 and Jeffrey Dean1 www.nature.com/npjdigitalmed Predicting in-hospital mortality: AUC-ROC of 0.93–0.94 30-day unplanned readmission: AUC-ROC of 0.75–0.76 Prolonged length of stay: AUC-ROC of 0.85–0.86 Final discharge diagnoses: AUC-ROC of 0.90 ిࢠΧϧςͷ࣌ܥྻ৘ใશ෦͔Βɺױऀͷ༧ޙ༧ଌ
95. ai.labo.ocu@gmail.com ਓ޻஌ೳݚڀࣨ AIʹؔ͢Δ૬ஊ͸͓ؾܰʹ