AI in medical imaging Flashcards
1
Q
Why is deep learning so successful?
A
End-to-end learning, learn features from data that are “optimal” for the given task
2
Q
Requirements for deep learning
A
Big data
Computational power
Algorithmic development
3
Q
Why Deep conv neural networks for images
A
- Images are high-dimensional input
- Huge number of weights for fully-connected networks
- Complicated to train
4
Q
filter size is independent of
A
the size of the image
5
Q
Options for MR acquisition
A
- Real-time MRI: fast, but 2D and relatively poor image quality
- Gated MRI: fine for periodic motion, e.g. respiration or cardiac motion but requires gating (ECG or
navigators) leading to long acquisition times (30-90 min).
6
Q
Deep learning for image reconstruction
A
Convolution + ReLU
Max pooling
Convolution + ReLU
Max pooling
Transposed convolution
Softmax
7
Q
Instability Test
A
- Tiny perturbations can lead to large artifacts.
8
Q
AI across the imaging pipeline
A
Disease Prediction
Disease Detection
Quantitative Markers
Semantic Interpretation
Image Enhancement
Image Reconstruction
9
Q
Image synthesis problem
A
High dependence on training set!
10
Q
which works better, late or early fusion?
A
late
