Nabz AI Academy - Watson Course - Session 1
Download as PPTX, PDF
0 likes142 views
悋 悴惺 悋愕悋惆 惡悋 惠惆惘愕 愆惠惘擧 愕惘擧悋惘 悽悋 悋愀 拆惘悋愕 惆惘 悴愕 悋 愕 惆惘 悋惠愕 悛擧悋惆 愆 惶惺 惡惷 悋惘悋悧 愆惆 悋愕惠
Nabz AI Academy - Watson Course - Session 1
- 1. Artificial Intelligence Basics in Healthcare Industry Fateme Pourghasem Sahand Samiei
- 2. Fateme Pourghasem Founder at BaMaMouz Health-IT Student at Tehran University of Medical Sciences Research Assistant at Company Non-Communicable Diseases Research Center (NCDRC) Entrepreneurship Mentor at National Scientific Olympiad for Medical Sciences
- 4. This person does not exist!
- 5. 悽悋悋惆擯 悋 悋 擧惆 悽悋悋惆擯 悋 悋 擧惆 愀悋惘 悋惘悋忰惆 0 惶悋忰 惶惆 1 慍悋惆 愀悋惡 忰悋惆 0 惺惡惆 悋愀 0 悋惆惘 惘悋 1 悋惆悋忰惆 悋愀 1 惘 悛惘 0 惶惘 愆惆惘悽 1 悴惺惘拆惘 惘 1 悋惆 愀 悋惘惘惷悋 0 悋悋 惺悋愀 0 惘愆 悽愆 悋惆悋悋 1 愕悋悋惘惺悋惡惆 惆 忰惆 1 悋忰惆慍悋惆 0 擧惆悽惆悋 擧悋 0 擧悽悋 悋悧慍 1 惘惺 擯悋慍 1 悋愕 愆悋擯悋 0 愕 忰惆 愕惆 0 拆惘悋愕 悋愀 惘 悋愕惆 惺 1 愕惺 愕惆 惴悋惘惠 悋愀 1 惡悋慍 !
- 6. 悴悋惡 惡悋慍 ! 悽悋悋惆擯 悋 悋 擧惆 悽悋悋惆擯 悋 悋 擧惆 愀悋惘 悋惘悋忰惆 0 惶悋忰 惶惆 1 慍悋惆 愀悋惡 忰悋惆 0 惺惡惆 悋愀 0 悋惆惘 惘悋 1 悋惆悋忰惆 悋愀 1 惘 悛惘 0 惶惘 愆惆惘悽 1 悴惺惘拆惘 惘 1 悋惆 愀 悋惘惘惷悋 0 悋悋 惺悋愀 0 惘愆 悽愆 悋惆悋悋 1 愕悋悋惘惺悋惡惆 惆 忰惆 1 悋忰惆慍悋惆 0 擧惆悽惆悋 擧悋 0 擧悽悋 悋悧慍 1 惘惺 擯悋慍 1 悋愕 愆悋擯悋 0 愕 忰惆 愕惆 0 拆惘悋愕 悋愀 0 惘 悋愕惆 惺 1 愕惺 愕惆 1 惴悋惘惠 悋愀 1
- 7. 惠惺惘 惶惺 愆
- 8. 惶惺 愆 惠惺惘 惠惺惘 : 惺悋 惡 惘悋 惶惺 愆 悋愆悋悽 擧 愆悽惶 擧悋拆惠惘 惺 悋慍 擧 惆 惡悋 悽惆擧悋惘愕悋慍 愆惆悋 惘惠悋惘悋 惆悋惘惆 愕惘擧悋惘 . 悋愕惠 悋 悋悴惘悋 愕悽惠 惡悽愆 : 惘悋 愆 悽惆 擧 悛悴悋 悋慍 惠悋 惆 惠惺惘 悋擧悋 擧 惠惺惘 惆惘愕惠 惡 惆悋惘惆 悴惆 惶惺 愆 .
- 9. 擧惘惆 惠愆惘忰 惡惘悋 惶惺 愆 悋惶愀悋忰 擧 愀惘 惡 悋愕愕惠 擧悋惘 惡 惘 惆 擧 惆 悋悛 悋慍 悋愕惠悋惆 悋悋愆 惠惆 惡惘悋 愕悋慍愆惡 惘惠悋惘悋 悋愕悋 愆 悛 惡悋 惘惠惡愀 悋愕惠 . 悋慍 悋愕惠悋惆 惡悋 悋愕惠 擧 擯悋 惆 悋 悋悋擯惘惠 愕悋惆 惠惺 拆愆 悋慍 悋擯悋 愆惆 惡 悋慍 擯悋 愆惆 忰 悋悋擯惘惠 悋惺悋惆 拆惆 惆悋惘惆 .
- 11. AI in healthcare
- 12. 悋惠擧 惘惆 悋愕惠悋惆 愆 愕悋惠 惆惘 惶惺 Machine Learning Machine Vision Natural Language Processing (NLP) Robotics
- 13. Disease Prediction : Traditionally approaches of doctors. Use of AI technologies in this area. Benefits.
- 14. X-RAY of a Human Hand
- 17. Surgery: da Vinci by Intuitive Surgical. Mako by Stryker. NAVIO by Smith & Nephew. Monarch by Auris Health.
- 18. 惘惡悋惠 愕悋 悴惘悋忰 ! 愕悋悋 悴惘悋忰 惘惡悋惠擧 愕悋 惡悽愆 惆 愆悋 悋惶 擧愕 悴惘悋忰 惆惘 惘悋 悋慍 悋惘惡悋惠 惡惘 愕惠惘 悴惘悋忰 惡悋 惡悋惘 悋愕惠 . 惘惡悋惠 悴惘悋忰 愕悋 惆悋惘悋 霸朸 悴慍 惶 惠惘 惡 悋擧惠惘擧 悋愕惠 悋惆惘 擧 悋愕惠 惠惘惴惘 悋惺 悴惘悋忰 惆惆 悋悴悋 惘悋 . 擧愕 悴惘悋忰 愆悋 悋惠惘 惆 惘惡悋惠 惘悋惡惘 拆惆悋悋 惆悋惠 慍惘拆悋 悴 擧 悋愕惠 惘悋忰 擧愕 拆愆惠 惆惘 愆愕惠 惡悋 惠惶悋惘 悋慍 愆惆 悋惘愕悋 悋忰 愆悋惆 惘悋 惺 悋惡慍悋惘悋 悴惘悋忰 惆惘惡 惠惶惘惡惘惆悋惘 惆惘 惘悋 悋慍 惶惘惠 惡 惘悋 惆悋惠 擧惆 . 悋慍 愕 惆 擯惘 愕 惘惡悋惠 拆惘 惆 愆悋 惘惡悋惠 悋惡慍悋惘 忰悋 擧 惘惡悋惠 惠惶惘惡惘惆悋惘 惆惘 擧 惡悋 惡 悋惘 愆惆悋惆 愕惠惘 悋悴惘悋 惆悋惘惆 惺惆 惡惘 惘悋 悴惘悋忰 惆愕惠惘悋惠 .
- 19. 慍悋悋 悋慍 惘惡悋惠 愕悋 悴惘悋忰 擧 悋慍 惠惘 擯悋 惘惡悋惠 悴惘悋忰 愕悋 悋慍 悋愕惠悋惆 悋惡慍悋惘悋 惠惆悋 悴惘悋忰 悋拆悋惘愕擧拆 惡 悴悋 惡 擧悋惘擯惘 悋惡慍悋惘悋 擧悋愆 擧 悋愕惠 悽悋惶 愆惆惆 慍悋 惶惘 惘惡悋惠 惆惘 惘悋 愀 惺 悴惘悋忰 惆悋惘惆 惆惡悋 惡 . 惡 愀惘 擧 慍 惠愕愀 惶惘 惘惡悋惠 愕悋 悋慍 悋愕惠悋惆 惶惘惠 惆惘 悋惡慍悋惘悋 惶惘 惆悋悧 惡愕悋惘 悋慍 惆惘 悋愕 惡悋 慍 惠愕愀 惘惡悋惠 惆悋 ( 忰惆惆 朶朧朧朧 惆悋惘 ) 悋愕惠 擧惠惘 惘悋惠惡 惡 .
- 21. 惶惺 愆 惠悋惘悽
- 24. 悋惘愆 忰惠 惘悋 惶惺 愆 惠悋 惺 惆惘 惆惘 惠悋惘悽 悋慍 惡愆惠惘 悋惺悋惠 : 束 悛 悧 損 Allen Newell 束 惘惡惘惠 悋 . 愕 損 Herbert A. Simon 束 悛 惠惘擯 損 Alan Turing 悴悴愕惠 擧惘惆 . 悛慍 愆惘 惠惘擯 惆惘 愕悋 霸杞杁朧 惠愕愀 悋 惆惘 悋悋 愀惘忰 愆惆 . 悋 悋 擧 悋慍 悋 悋愕悋惆 悋愕惠 擧 惆惘 悛 惡 悴惆 悛惆 悋悋愆 愆惆 惡拆愆 愆惆 悋愕惠 .
- 25. 悴悋 惠愕愀 惶惺 愆 愃惠 悛惆 悴惆 惡 擧悋惘惠擧 愕悋 惆惘 霸杞杁杁
- 28. 悋惠愕 - 朮朧霸霸
- 29. AlphaGo 愕悋 朮朧霸朷
- 31. 悽惠 悋悋惺 惶惺 愆
- 32. 悋愆悋悽 擯悋擯 惆惘 惶惺 愆 悋慍 悋惆悋愆 悋惘悋悋 悋愕惠悋惆 惘惆 惘悋惘 擯惘惆 悋 惡惘悽 悋愆悋悽 悋慍 惺惡悋惘惠惆 : 悋愆 悋惆擯惘 Machine Learning 惶惺 惺惶惡 愆惡擧 Neural Networks 悋愆 惡悋 Machine Vision 悋愕悋悋 悽惡惘 Expert System 愀惡惺 慍惡悋 拆惘惆悋慍愆 NLP 悋擯惘惠 惠擧 Genetic Algorithm 惡悋 惘惠惡愀 悋 惘惡悋惠擧 Robotic
- 33. 悋愆惡擧 惺惶惡
- 34. 忰惆悋 惡悋 悋愀悋惺悋惠 惆惘悋惠 悋慍 拆愕 惘悛惆 悋 惆 悋 惆擯惘 ( 惆惘 悋愕愕惠 惡慍惘擯 惡愆 惠悋 悋慍 惡愕惠 悋 ) 拆惆悋 悋惆悋 擧惆 惶惘惠 惡 擧 愕愕 惘悋惠惡 惆悋惘惆 惘悋惘 惘 拆惆悋 悋慍 悋愕惠悋惆 惡悋 惘悋 悋愀悋惺悋惠 惡惆惆愕惠 悋惘愕悋 擧惆 . 惆惘 惠惘拆悋 惡悽 愕愕 愆 惘悋惠惡 悋 惠惺惆悋惆 惺 愀惘 惡 擧 惆悋惘惆 惘悋惘 悽惘悴 悋惘 惠悋 悋慍 悛 惶惺 悋悋 悋愕惠 擧惠惘 惆擯惘 . 悋 悋 悋惆悋惆 惘悋 愆惆 忰悋愕惡 惡 惘惠 惠惡惆 擧惆 擧 惡惘悋 惡悋愆惆 悽悋惆 悋惡 悋愆 .
- 35. 惓悋 : 悽悋 惠
- 37. Supervised
- 38. Unsupervised
- 39. Reinforcement
- 40. Email spam? (0/1) spam filtering Audio text transcript speech recognition English Persian machine translation Image, raider info position of other cars self-driving car Image of phone defect? (0/1) visual inspection Input (A) output (B) Application
- 42. Classification
- 43. 惺 惆 supervised :
- 44. 惺 悋惆擯惘 ( Deep Learning ) 悋慍惘擯 悋愆 悋惆擯惘 悋慍 悋 悋 惺惶惡 惶惺 愆 悋慍 惘惆 悋 惆惘 悋愆惡擧 悋愕惠悋惆 惺惶惡 愆惆 . 悋擧惓惘 惡惘 惶惺 愆 悋惘慍 擧悋惘惡惘惆悋 拆悋 愕惠 惺 悋惆擯惘 惆 .
- 48. 悋悋愆 愕悋惠 惆惘 惶惺 愆 Training Doctors/Patients Adoption Regulations Maintenance Security Data!
- 49. 5 minutes break
- 50. Sahand Samiei Medical Advisor at Nabz Group, AI Team Medical Extern at Tehran University of Medical Sciences Innovation Administrator at TUMS Exceptional Talents Development Center
- 51. Where are we?! Job Positions Domain Expert or Interdisciplinary Programmer Common Literature & Coordination Problem-Solution Fit Assessment Generalism
- 52. A.I. top benefits in health industry Systematic collection of big data & Information accessibility Careful decisions with high accuracy Streamlining processes with high speed Scalability of health services Reduce costs
- 53. Ethical concerns Bias Data collection Automation
- 54. an ML example for diagnosis Mass Normal Mass ?!
- 55. an ML example for diagnosis Image Mass Probability 0.48 0.51 Desired Label: 1 (Mass) Desired Label: 0 (Normal) Error (Loss) = 0.32 Error (Loss) = 0.31 Image Mass Probability 0.60 0.30 Desired Label: 1 (Mass) Desired Label: 0 (Normal) Error (Loss) = 0.22 Error (Loss) = 0.15
- 56. Class Imbalance Challenge Mass Normal Normal Normal
- 57. Binary Cross Entropy Loss Function Algorithm Algorithm 0.2 Label 1 0.7 Label 0
- 58. Impact of Class Imbalance on Loss Calculation Weighted Loss Loss Prediction Probabilities Examples 2/8 x 0.3 = 0.075 0.3 0.5 P1 Normal 2/8 x 0.3 = 0.075 0.3 0.5 P2 Normal 2/8 x 0.3 = 0.075 0.3 0.5 P3 Normal 6/8 x 0.3 = 0.225 0.3 0.5 P4 Mass 2/8 x 0.3 = 0.075 0.3 0.5 P5 Normal 2/8 x 0.3 = 0.075 0.3 0.5 P6 Normal 6/8 x 0.3 = 0.225 0.3 0.5 P7 Mass 2/8 x 0.3 = 0.075 0.3 0.5 P8 Normal
- 59. Resampling to Achieve Balanced Classes Examples P1 Normal P2 Normal P3 Normal P4 Mass P5 Normal P6 Normal P7 Mass P8 Normal Loss Prediction Probabilities Re-Sampled 0.3 0.5 P3 Normal 0.3 0.5 P6 Normal 0.3 0.5 P1 Normal 0.3 0.5 P8 Normal 0.3 0.5 P7 Mass 0.3 0.5 P4 Mass 0.3 0.5 P7 Mass 0.3 0.5 P4 Mass
- 60. Multi-Task Mass ?! Pneumonia ?! Edema ?! Algorithm Binary Task Algorithm Binary Task Algorithm Binary Task Algorithm Binary Task Algorithm Multi Task
- 61. Multi-Task Weighted Loss Prediction Probabilities Examples 0.52 + 1.00 + 0.70 0.3, 0.1, 0.8 P1 0, 1, 0 0.05 + 0.05 + 0.10 0.1, 0.1, 0.8 P2 0, 0, 1 0.10 + 0.70 + 0.15 0.2, 0.2, 0.7 P3 0, 1, 1 0.22 + 0.52 + 0.10 0.6, 0.3, 0.8 P4 1, 0, 1 0.15 + 0.15 + 0.05 0.7, 0.7, 0.9 P5 1, 1, 1 0.10 + 0.05 + 0.10 0.8, 0.1, 0.2 P6 1, 0, 0 0.52 + 0.05 + 0.10 0.3, 0.9, 0.8 P7 0, 1, 1 0.05 + 0.05 + 0.10 0.1, 0.1, 0.2 P8 0, 0, 0
- 62. Dataset Size ~ 10K Samples ~ 100K Samples ... Solutions 1. Pretraining + Fine Tuning (Transfer Learning) 2. Data Augmentation CNN Penguin / Cat / Dog CNN Mass / Pneumonia / Edema
- 63. Key Challenges Multi Label Loss Weighted Loss / Resampling Transfer Learning / Data Augmentation Multi-Task Class Imbalance Dataset Size
- 64. Model Testing Dataset Training Set Validation Set Test Set Development of models Tuning and selection of models Reporting of results Cross validation
- 65. Patient Overlap Dataset Split by image Training Set Validation Set Test Set Mass P. ID Image ... 1 20 CXR1.JPg ... 0 17 CXR4.JPG ... 0 32 CXR7.JPG ... ... ... ... Mass P. ID Image ... 1 20 CXR2.JPg ... 0 11 CXR5.JPG ... 0 32 CXR9.JPG ... ... ... ... Mass P. ID Image ... 1 20 CXR0.JPg ... 0 38 CXR3.JPG ... 0 32 CXR8.JPG ... ... ... ...
- 66. Patient Overlap Dataset Split by patient Training Set Validation Set Test Set Mass P. ID Image ... 1 20 CXR1.JPg ... 1 20 CXR2.JPG ... 1 20 CXR0.JPG ... ... ... ... Mass P. ID Image ... 0 32 CXR7.JPg ... 0 32 CXR8.JPG ... 0 32 CXR9.JPG ... ... ... ... Mass P. ID Image ... 0 17 CXR4.JPg ... 0 38 CXR3.JPG ... 0 11 CXR5.JPG ... ... ... ...
- 67. Set Sampling 10% of data ---------------- 120 CT Scans 400-500 X-Rays 130 Whole 際際滷s X% of minority class !
- 68. Ground Truth (Reference Standard) Consensus voting 1) + Yes Yes No Mass ?? = Yes 2) + More definitive test
- 69. Key Challenges Minority class sampling Split by patient Consensus voting / More definitive test Set Sampling Patient Overlap Ground Truth
- 70. Accuracy = Examples correctly classified Total number of examples Accuracy = P(correct) Accuracy = P(correct disease) + P(correct normal) Using P(AB) = P(A|B) P(B) : Accuracy = P(correct|disease) P(disease) + P(correct|normal) P(normal) Accuracy = P( + |disease) P(disease) + P( - |normal) P(normal) Evaluation Metrics Model 2 Model 1 Ground Truth - - Normal + - Normal - - Normal - - Normal - - Normal + - Disease - - Normal + - Disease + - Normal - - Normal Sensitivity (True positive rate) Specificity (True negative rate)
- 71. Evaluation Metrics TP FP TN FN - + Disease Normal Accuracy = P(correct) Accuracy = Sensitivity x P(disease) + Specificity x P(normal) Prevalence Accuracy = Sensitivity x Prevalence + Specificity x (1 Prevalence) 1. Sensitivity = TP / (TP + FN) 2. Specificity = TN / (FP + TN) 3. PPV = TP / (TP + FP) 4. NPV = TN / (TN + FN)
- 72. ROC curve & Threshold Model Probability > t ? + - Output Probability (Score) X-Ray 0.30 1 0.42 2 0.78 3 0.98 15 0 Score 1 Negative t Positive
- 73. Resources for further study
- 74. Resources for further study
- 75. The greatest enemy of knowledge is not ignorance, it is the illusion of knowledge! - Stephen Hawking
- 76. Q&A