Inhoudsopgave
LC 2.1 AI: basics of artificial intelligence in medicine..............................................................................1
LC 2.2 AI in diagnostic pathology............................................................................................................1
PC2 Basics of AI.......................................................................................................................................3
LC 2.3 AI in skin tumor............................................................................................................................4
LC 2.4 AI impact reader performance.....................................................................................................5
LC 2.1 AI: basics of artificial intelligence in medicine
AI first made completely by humans (experts) guided by knowledge. Then Machine learning (ML) also
partially designed by computers (when they became more powerful) build on statistics/examples.
Lately booming is the Deep Learning (DL) which is almost completely made by computers, using
multiple layers to extract high-level features from the raw input.
LC 2.2 AI in diagnostic pathology
How to build an AI system:
In the past: make up important but simple features like colors, cell count which you give to a
machine. Classification in a graph to make up the groups cancer or not. However, this is too simple
for harder pictures to define.
, Nowadays: CAD system with neural network as inspiration. Signals are added together. Activation
function determines which will be put as output.
Deep neutral networks are networks of artificial neurons, which have input (each input is a weight)
and if the weights together crosses a threshold an output is given. The layers in between the input
and output layers are hidden layers.
Hidden layer output for every neuron = the amount of inputs + bias (1).
Capacity of neural network = more parameters (input to neurons) the more complex
Building an AI:
1. Define an architecture = how many layers and neurons. Network architecture = amount of
layers and neurons
2. Initialization = giving all parameters/neurons a value (random initialization)
3. Training and validation set = train / optimize the network with a training set. Asses its
performance with validation set (monitor training loss and accuracy -> accuracy goes up and
error goes down)
Overfitting = not learning but memorizing the training set -> makes more mistakes in other images.
Convolutional neural networks are different from normal neural networks. They are based on our
visual networks in the brain (response to horizontal and not vertical e.g.), so filtering instead of
multiplying. Certain neurons in the brain respond to very specific stimuli. Used with medical images
(more efficient with parameters)
Convolution = image or signals and random initial filter. Put the filter on the image and fill the matrix
with found data. In a convolutional neural network the filters are learned/specialized to predict the
image/data input. The values of the filters are updated to reduce the error (training).
- A convolution = operation with an image/signal with a filter (smaller image) you move across
the image. At every location it fits you multiply it and create a new image. If the filter
matches exactly with the image (purple circle) it will give the highest response
- Downsample = make the image twice as small.
Using these neural networks for tasks in medical imaging:
We can divide the usage of deep learning models in three categories in our field:
- Object detection: classification of every point in the image (is not very efficient) (find amount
of specific mitotic cells for example) and determining the location of the mitosis = post-
processing
- Segmentation: classification for predicting cancer by dividing image in regions every region is
an object (classify what pixel belonging to tumor and what to stroma for example)
- Classification (We show the model an image and it gives out the label, for example malignant
tumor, or not. = end to end processing
LC 2.1 AI: basics of artificial intelligence in medicine..............................................................................1
LC 2.2 AI in diagnostic pathology............................................................................................................1
PC2 Basics of AI.......................................................................................................................................3
LC 2.3 AI in skin tumor............................................................................................................................4
LC 2.4 AI impact reader performance.....................................................................................................5
LC 2.1 AI: basics of artificial intelligence in medicine
AI first made completely by humans (experts) guided by knowledge. Then Machine learning (ML) also
partially designed by computers (when they became more powerful) build on statistics/examples.
Lately booming is the Deep Learning (DL) which is almost completely made by computers, using
multiple layers to extract high-level features from the raw input.
LC 2.2 AI in diagnostic pathology
How to build an AI system:
In the past: make up important but simple features like colors, cell count which you give to a
machine. Classification in a graph to make up the groups cancer or not. However, this is too simple
for harder pictures to define.
, Nowadays: CAD system with neural network as inspiration. Signals are added together. Activation
function determines which will be put as output.
Deep neutral networks are networks of artificial neurons, which have input (each input is a weight)
and if the weights together crosses a threshold an output is given. The layers in between the input
and output layers are hidden layers.
Hidden layer output for every neuron = the amount of inputs + bias (1).
Capacity of neural network = more parameters (input to neurons) the more complex
Building an AI:
1. Define an architecture = how many layers and neurons. Network architecture = amount of
layers and neurons
2. Initialization = giving all parameters/neurons a value (random initialization)
3. Training and validation set = train / optimize the network with a training set. Asses its
performance with validation set (monitor training loss and accuracy -> accuracy goes up and
error goes down)
Overfitting = not learning but memorizing the training set -> makes more mistakes in other images.
Convolutional neural networks are different from normal neural networks. They are based on our
visual networks in the brain (response to horizontal and not vertical e.g.), so filtering instead of
multiplying. Certain neurons in the brain respond to very specific stimuli. Used with medical images
(more efficient with parameters)
Convolution = image or signals and random initial filter. Put the filter on the image and fill the matrix
with found data. In a convolutional neural network the filters are learned/specialized to predict the
image/data input. The values of the filters are updated to reduce the error (training).
- A convolution = operation with an image/signal with a filter (smaller image) you move across
the image. At every location it fits you multiply it and create a new image. If the filter
matches exactly with the image (purple circle) it will give the highest response
- Downsample = make the image twice as small.
Using these neural networks for tasks in medical imaging:
We can divide the usage of deep learning models in three categories in our field:
- Object detection: classification of every point in the image (is not very efficient) (find amount
of specific mitotic cells for example) and determining the location of the mitosis = post-
processing
- Segmentation: classification for predicting cancer by dividing image in regions every region is
an object (classify what pixel belonging to tumor and what to stroma for example)
- Classification (We show the model an image and it gives out the label, for example malignant
tumor, or not. = end to end processing
