what questions can we ask?
- descriptive
- diagnostic (why?)
- predictive (what will happen?)
- perspective (what should we do?)
3 case studies
arrhythmia detection in important. electrical activities of heart. Zio patch (up to 2 weeks = 1.6 billion heart beep)
the burst of data will help automatic detection aps grow
difficult to diagnose with single lead ECG (normally we do that with 12 leads)
differences are quite subtle in heart arrhythmia
1d conv net over time dimension of input
residual nets(short cuts for very deep net)
64 k ECG records (600x bigger than MIT-BIH)
surpass human level by 3 percent !
continuously monitor patient
apple watch
detect pneumonia from chest x-rays pictures
2 billion per year images
- 2d CNN 224*224
- pretrained on image net
- 121 layer dense net
dense net : connect all of the layers together (all have shortcut to each other)
112,120 frontal view x-ray 30k patient (largest Sep. 2017)
NLP systems that reads reports
420 test set with Stanford x-ray expert
we don't have ground truth so we check them with each other and compute f1 score
Chex-net (435 and expert 395)
Trust is an issue. class activation maps = ways of look at what part of images are imp to model.
benefits of model
- improve healthcare delivery
- increase access to MRI
multi label problem
- abnormality
- ACL
- Meniscule
1370 MRI exams
- abnormality = 0.937
- ACL = 0.965
- Meniscule =0.847
external validation
different scanner different country. and it did relatively well without training.
MURA = largest dataset for abnormality detection
determine microscopic image correspond to image
segment images = classification of each pixels.
we should have an mask at output.
100 k images:
-
how to distribute this if C is the important data?
90 - 5 -5 where 5 - 5 all from C and some C are in Train Data
-
image augmentation? crop - scale - blur - zoom - translate - H flip - random noise - changing contrast - etc.
-
can image augmentation hurt an algorithm?
Yes, in case of character recognition.
(little case study: problem of car speech recognition when driving backward? mic was on the front and the sound was vague and they augmented vague data)
We label like this (Cell , background, boundary)
for each use softmax of pix classes of y - > (0,0,1)
class unbalance will happen. and we use class weight ratio
binary classification . how to state why the model chose this answer? derivative of y^ in respect to x will show you a matrix of shape x. this matrix will have higher numbers in areas witch exchanging it's pixels will change the model's prediction. and lower numbers in areas witch changing it's pixels are making no difference.
q : can we make 99% accuracy and human level performance is 97%. is it possible?
yes if we are competing with one doc and our knowledge is coming from a group of docs.