Using Artificial Intelligence in the field of Diagnostics_Case Studies.ppt
- 1. Using Artificial Intelligence in the field of Diagnostics
- 2. 2,500,000,000 Gigabytes of data produced daily 2
- 3. 80% Percent of World’s Population in low-income communities. 3
- 4. ▫ Large amount of data perfect for AI ▫ Room for improvement ▫ Millions of lives affected by improvements in this field AI IN DIAGNOSTICS 4
- 5. What is a Neural Network? 5
- 6. AI for Diabetic Retinopathy Eyeagnosis Teammates: • Neeyanth Kopparapu • Kavya Kopparapu • Justin Zhang
- 7. 332,000,000 Diabetic patients in middle-low income communities 7
- 8. > 50% Estimated cases of diabetes still undiagnosed 8
- 9. 9
- 10. Diagnosis for Diabetic Retinopathy needs to be: - Inexpensive - Quick - Accessible - (finally) Accurate PROBLEM 10
- 11. 11
- 12. Efficient • Subjectivity • Requires >2 hours/patient • Ophthalmologist : Patient > 1:1000 USING ARTIFICIAL INTELLIGENCE Inexpensive • Current diagnostic tools are expensive • Pathologist required to diagnose 12
- 13. ▫ Accessibility use a common devices ▫ Inexpensive don’t use expensive equipment ▫ Quick AI models are very quick ▫ Diabetic Retinopathy common procedures include dilated eye exam, fluorescein angiography, optical tomography DRAFTING A SOLUTION 13
- 14. SOLUTION 14
- 15. NEXT STEPS Assemble Dataset NIH eyeGENE Dataset of retinal images Train Model ResNet-50 Convolutional Neural Network Create Supporting Technology - Mobile Application - Lens Holder 15
- 16. What is a Convolutional Neural Network? 16
- 17. DATASET 17
- 19. MODEL RESULTS 19
- 20. 20 SUPPORTING TECHNOLOGY - Mobile Application - Facilitate Transfer of Image - Preprocess Image - Lens System - Holds multiple lenses - Better picture of the Retina
- 22. LENS SYSTEM 22
- 23. Can We Do Better?
- 24. 24 HOW CAN WE IMPROVE? - Accessibility - Accuracy - Trust
- 27. AI for Major Depressive Disorder TweePression teammates: • Neeyanth Kopparapu • Kaien Yang
- 28. 41.6% Percent of Students with Depression 28
- 29. 43% Percent of Students with Depression that seek help 29
- 30. Accurate • Current depression screening methods have accuracies under 75% A DIFFERENT PROBLEM Non-Intrusive • Over 40% of adults keep depression symptoms to themselves • Opening up to doctors is harder than online (Social Media) 30
- 31. DRAFTING A SOLUTION ▫ Accuracy Bigger collection of data, more viewpoints ▫ Non-Intrusive Social Media? 31
- 32. USING ARTIFICIAL INTELLIGENCE Sentiment Analysis From individual social media posts, calculate sentiment of user Depression Prognosis From Sentiment Analysis, give final prediction about user. 32
- 33. 33 INITIAL APPROACH ▫ Classify each word individually ▫ Combine classifications for final output
- 34. RESULTS 34 True Positive 104,460 True Negative 121,065 False Negative 11,242 (Type II Error) False Positive 13,233 (Type I Error) Condition Positive Condition Negative Test Positive Test Negative
- 35. Can We Do Better?
- 36. • The main goal of using AI is to increase accuracy • Using as much data as possible is beneficial • Accessibility isn’t the immediate problem • Using medical machines to increase data output REANALYZING THE PROBLEM 36
- 37. BAG OF WORDS ▫ “How could I hate you?” ▫ “He said he loved ONLY me” vs “He said ONLY he loved me” 37 How Could I You? Hate
- 38. NEW MODEL 38 2: CONVOLUTIONAL LAYER Sentiment Analysis Dataset Kaggle Movie Reviews Dataset Google Word2Vec DENSE NEURAL NETWORK 1: EMBEDDING LAYER 3: MAX POOLING 2D LAYER OUTPUT: SOFTMAX ACTIVATION 5: DENSE LAYER 4: DENSE LAYER
- 39. RESULTS 39 True Positive 240,303 True Negative 184,257 False Negative 9,702 (Type II Error) False Positive 15,738 (Type I Error) Condition Positive Condition Negative Test Positive Test Negative
- 40. FMRI ANALYSIS MODEL 40 Convolution + ReLU Max pooling Fully Connected + ReLU Softmax Image Input: fMRI Scans Classification Output: MDD Prediction
- 41. FMRI RESULTS 41 True Positive 446 True Negative 372 False Negative 29 (Type II Error) False Positive 28 (Type I Error) Condition Positive Condition Negative Test Positive Test Negative
- 42. SUPPORTING TECHNOLOGY ▫ Server to process queries - Validation of Social Media, Runs AI Model, Outputs Result ▫ Website for user interface with access to social media pages 42
- 44. Looking Forward
- 45. AI IN DIAGNOSTICS ▫ Will it Replace Doctors? - Loss of Jobs? ▫ How can we justify outputs? - Trust among patients? ▫ How can we scale this? ▫ What should be prioritized? ▫ What happens if the AI is wrong? 45
- 46. 46 CURRENT PROBLEMS IN AI ▫ Biased datasets ▫ Unethical training procedures ▫ Inaccessibility to data
- 47. 47 FINAL THOUGHTS ▫ Diagnostics is the perfect field for AI ▫ However, there is change that must come before Explanations ▫ Adaption toward specific communities
Editor's Notes
- #3: Say the Statistic, then say “5 years ago our current state of technology could not keep up… this is really where artificial intelligence began to rise”
- #4: Stay statistic… “More specifically, AI has started to rise in the field of medicine and healthcare because there are many people that can’t access the expensive healthcare systems in many countries today”
- #5: “AI has a lot of potential in diagnostics since the modern data revolution… each medical device from MRI scanners to Glucose Meters record data to send to a doctor or to store for a patient. AI can capitalize on this information to form predictions.
- #6: “In general, NN’s in the most basic sense are just layers of neurons that are connected with neurons in other layers. To send information through the neuron, the signal is passed through a neuron, send through a nonlinear function, and if the final number exceeds a threshold, it fires to all of its connections, until it reaches the output layer, whose information is passed to the user… Backpropogation involves adjusting weights to fit information that was passed better. Talk about what happens for the rest of the presentation.
- #7: First application I worked on was using AI for Diabetic Retinopathy
- #10: Explain what Diabetic Retinopathy is a complication of diabetes that affects the eye, caused by damage to blood vessels, and poorly controlled blood sugar.
- #12: DR Is RIGHT?? LOOKS IDENTICAL!
- #14: Dialiated Eye Exam Eyedrops and then Retinal Image Fluorescein Angiography Using injected dye to pinpoint broken & closed blood vessels. Optical tomography shows thickness of retina.