6. 의료 AI를 시작하기 전 제가 알던 모습…
[1] Predicting non-small cell lung cancer prognosis by fully automated microscopic pathology image features, http://www.nature.com/articles/ncomms12474
[2] SURVIVALNET: PREDICTING PATIENT SURVIVAL FROM DIFFUSION WEIGHTED MAGNETIC RESONANCE IMAGES USING CASCADED FULLY CONVOLUTIONAL AND 3D
CONVOLUTIONAL NEURAL NETWORKS, https://arxiv.org/pdf/1702.05941v1.pdf
8. Deep learning is Data-Driven approach
MORE DATA
[1] Revisiting Unreasonable Effectiveness of Data in Deep Learning Era, https://arxiv.org/pdf/1707.02968v1.pdf
9. Deep learning is Data-Driven approach
MORE DATA
REAL WORLD
Training
data
10. Data is problem !!
MORE DATA, GOOD DATA
REAL WORLD
Training
data
18. 의료 데이터 문제 이유?
1. 의사는 학습 데이터 만드는 사람이 아니다!
2. 의료 데이터는 바둑 기보가 아니다!
3. 병원은 환자의 데이터가 있는 곳이다!
4. 의료 데이터는 일반 데이터와는 그 성질이 다르다!
19. 데이터 문제 해결을 위한 시도들?
AI with Data vs. AI for Data
20. Data, more data
Deep Learning is Robust to Massive Label Noise, https://arxiv.org/pdf/1705.10694v2.pdf
21. Data, more data
Deep Learning is Robust to Massive Label Noise, https://arxiv.org/pdf/1705.10694v2.pdf
22. Pre-trained Model with adequate fine-tuning
Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning?, https://arxiv.org/abs/1706.00712
e use of the fine-tuning approach > full training from scratch
23. Pre-trained Model with adequate fine-tuning
Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning?, https://arxiv.org/abs/1706.00712
28. Data selection using Uncertainty
Leveraging uncertainty information fromdeep neural networks for disease, detectionhttp://biorxiv.org/content/biorxiv/early/2016/10/28/084210.full.pdf
29. Reverse Classification
Reverse Classification Accuracy: Predicting Segmentation Performance in the Absence of Ground Truth (http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7902121),
IEEE TRANSACTIONS ON MEDICAL IMAGING, 2017)
30. Medical Image Synthesis
Medical Image Synthesis with Context-Aware Generative Adversarial Networks (https://arxiv.org/pdf/1612.05362v1.pdf)
34. Publish Houses of Brick,
not mansions of straw
https://www.nature.com/news/publish-houses-of-brick-not-mansions-of-straw-1.22029
The main question when
reviewing a paper should be
whether its conclusions are likely
to be correct, not whether it would
be important if it were true. Real
advances are built with bricks, not
straw.
35. 의료 데이터 문제 해결을 위해
1. 일반 AI에서 쓰인 문제 해결 전략도 의료 데이터에 쓰일
수 있음!
2. 의료 데이터에 맞게 변형된 전략이 사용될 때 좋은 성능
보임!
그러나 많은 전략들은
‘문제의 해결’이 아닌 ‘문제의 완화’
의료 AI는 문제의 해결을 위해 계속 노력해야 하는 분야
36. 한국 의료 AI는 어떻게 문제를 해결해야 하는가?
AI or 의료 vs. 의료 with AI: 최고의 스승이자 고객
37. If AI kids do not know medical….
Virtual PET Images from CT Data Using Deep Convolutional Networks: Initial Results , https://arxiv.org/pdf/1707.09585.pdf
38. Pictures Are Not Taken in a Vacuum
ctures are not taken in a vacuum – An overview of exploiting context for semantic scene content understanding, IEEE Signal Processing Magazine ·April 2006
39. Medical image + context data
Prediction of Kidney Function from Biopsy Images Using Convolutional Neural Networks, https://arxiv.org/pdf/1702.01816v1.pdf
Training with initial eGFR:
Decreasing the training time by 2x
Decreasing the validation error by 20%
40. Medical image + context data
A Multi-Armed Bandit to Smartly Select a Training Set from Big Medical Data (https://arxiv.org/pdf/1705.08111v2.pdf), MICCAI 2017
41. Medical image + Medical image
Brain Tumor Segmentation with Deep Neural Networks, https://arxiv.org/abs/1505.03540
42. The efficient way to use context data ?
Real-Time User-Guided Image Colorization with Learned Deep Priors (https://arxiv.org/abs/1705.02999, SIGGRAPH 2017)
43. The efficient way to use context data ?
What Can Help Pedestrian Detection? (CVPR 2017, https://arxiv.org/pdf/1705.02757v1.pdf)
44. The efficient way to use context data ?
Image Question Answering using Convolutional Neural Network with Dynamic Parameter Prediction, https://arxiv.org/pdf/1511.05756.pdf
45. The efficient way to use context data ?
A simple neural network module for relational reasoning, https://arxiv.org/pdf/1706.01427.pdf
46. 의료 데이터는 환자의 데이터!!
1. AI 전문가만으로의 연구는 한계가 분명!
2. 기존 의학 지식/수치등은 AI의 좋은 스승이자 데이터!
3. Deep learning은 이질적인 데이터를 융합하기에 적합!
의료 AI는
‘AI 전문가가 데이터만 있으면 해결할 수 있는 문제’가 아니
라,
‘의료 전문가와 AI 전문가가 경험과 데이터를 가지고
함께 풀어야 하는 문제’!!