Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5525_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
31.08.2026
Размер:
29 Мб
Скачать
Artificial Intelligence in Adaptive Radiation Therapy
9.2.2 What is deep learning?
Deep learning can be dened as a subset within machine learning. Whereas machine learning combines several features directly into an output, deeplearning is often associated as having several intermediary or hiddenlayers. A single hidden layer, as shown in gure 9.1, would most appropriately be referred to as a shallow network, with deeprequiring two or more hidden layers. The benet of multiple hidden layers is the ability to represent predictions beyond simple linear regression when non-linear activation functions are used (activation functions are spoken of more in the following sections). The downside to the increase in the prediction capabilities of the model is the decrease in interpretability of the model predictions. Deep learning is often cited as being a black box, although this does not mean it is impossible to crack open the box and take a peek inside now and again.

9.3 Deep learning: the basic components

While deep learning has evolved considerably over the past decades, most models can be broken down into a combination of several key components. Having a solid understanding of each part will help one understand better how more complex models are generated in the future.
9.3.1 Convolutional neural networks: looking at the picture
Convolutional layers form the bedrock of deep learning image classication, segmentation, and registration models. The basis of these three tasks all relies on the same thing: the computer needs to understand what is occurring within an image. Humans, as well as most animals, have the benet of eyes which enable us to capture a large amount of information within our eld of view. For a computer this can be a relatively daunting task.
The rst thing which needs to be dened is the computers receptive eld of view. A convolution is very simply the creation of a new matrix by multiplication and resultant addition of elements on an image. A visual representation of this concept is shown in gure 9.2. Here, the input image is convolved with a simple three-by-three kernel to create a resultant feature image.
If the kernel were a vertical line surrounded by zeros, the output feature map would show where vertical lines are present in the image.
One of the downsides of convolutions is that they are locally dependent. This means that since a kernel size is xed, images of varying zoom or scale might not properly align with a previously dened kernel. A natural question which arises from this, is why not push towards large kernels which are able to cover large portions of the input image? There are two main concerns with this approach: (i) the larger the kernel, the more computational resources are required to perform the convolution and addition and (ii) larger kernels have more variables present, where an n by m matrix will have n × m + 1 variables. A three-by-three kernel has nine variables (plus a bias to make ten total), while a nine-by-nine kernel has 81 variables (plus a bias to make 82).
9-4
Artificial Intelligence in Adaptive Radiation Therapy
Figure 9.2. General representation of a convolution. The input image (left) convolved with the 3 × 3 kernel (middle) creates a feature map (right).
Figure 9.3. Example of a two-by-two max-pooling layer.
This modest increase in kernel size creates significantly more variables to train. A remedy to this local dependency issue comes in the form of pooling layers.
9.3.2 Pooling layers: keeping what matters most
Pooling layers are a way of taking only the most important aspect of a generated feature map. While they come in a variety of forms, a very commonly used pooling layer is max-pooling. Pooling layers are like convolutional layers in that they are a matrix, typically of size two-by-two for two-dimensional images, but rather than multiplying and adding, a max-pooling layer will simply take the maximum value within the dened matrix size, gure 9.3.
One of the most immediate implications of this is that the feature map itself has now decreased in size by a factor of 2 in each direction. This can either be seen as benecial or detrimental, depending on the desired output. From a positive perspective, max-pooling layers have multiple benets. The rst of these benets is that the most intense, or important, feature is maintained, while presumably less important features are removed. Second, subsequent convolutions on the
9-5
Artificial Intelligence in Adaptive Radiation Therapy
post-max-pooled image require signicantly less memoryas stated before, the feature map has been reduced in size. A 2 × 2 max-pooled image will be 1/4 the original image size, meaning less memory is required for the convolutional operation. Third, subsequent convolutions now cover a relative area that is larger than the original image size. A three-by-three kernel on an image with resolution of 1 mm × 1 mm per voxel would have a receptive eld size of 3 mm by 3 mm. After max-pooling, the resolution of the image is 2 mm × 2 mm per voxel, so a three-by­three kernel has a relative receptive eld of view of 6 mm by 6 mm. This is particularly useful in images where the scale/zoom of the image can vary. If the model is designed to identify that a dog is present within an image, it should not matter how large or small the dog presented is.
Despite these benets for image classication, if the user wishes to identify individual voxels relating to a class (segmentation), the pooling layers can lead to a signicant decrease in resolution. A representation of max-pooling followed by bilinear up sampling to the original image size is shown in gure 9.4. Note how the ne resolution text is quickly lost by the fourth pooling, while general features such sa color can be maintained for multiple layers. Maintaining the higher-resolution information for later decision making is a main factor in the popularity of the U-Net style architecture discussed later in this chapter.
One thing we have yet to cover is how these features can lead to a prediction. A model needs a method of combining these features into a nal prediction; this combination of features is dened as a fully connected layer.
Figure 9.4. Illustration of multiple max-pooling layers and bilinear resampling to illustrate how ne resolution information (text) is quickly lost, while large information (general colors) remains.
9-6
)
Artificial Intelligence in Adaptive Radiation Therapy
9.3.3 Fully connected (dense) layers: bringing it all together
Fully connected, or dense, layers can be imaged as the opposites of convolutional layers. While convolutional layers are locally dependent, the fully connected layer will combine every aspect of the input into an output, gure 9.5. The number of inputs and outputs in a dense layer can vary and are user dened. The mathematical value of each output is
out
i
out
+
Fh W B
(
i
i
1
ii
, where F is an activation function,
W is the weighting matrix, and B is the bias.
Within convolutional neural networks there can be any number of convolutions and pooling layers. These maps can be represented as several n by m matrices, where n and m are likely smaller than the original image dimensions. In order to feed these features into a dense layer they must rst be attened. This can be imaged as taking the feature image and laying it into a single vector: converting the n-by-m matrix into an n × m vector. These features can then be used in the nal classication. For the example in gure 9.6, we might say that if ears, whiskers, and paws are present, then a dog is present in the image.
The low-level and abstract features of lines, curves, or fuzziness are most likely to appear early in the architecture. Higher level features, such as ears and whiskers, can be imagined arising deeper, as a combination of overlapping low-level features. Finally, all of these features can be combined together after attening to identify if a dog is present somewhere in the image. Note that our output does nothing to identify where the dog is present. Furthermore, if the model is truly identifying whether ears, whiskers, and paws are present, it very likely confuses a cat for a dog.
Figure 9.5. A fully connected, or denselayer. Output features are a function of weights from each input feature.
9-7
Artificial Intelligence in Adaptive Radiation Therapy
Figure 9.6. Generalized example of a convolutional neural network for the identication of a dog. The rst convolutional layers are very low-level; lines, edges, fuzziness. Further down, these features combine to identify ears, paws, or whiskers, before nally feeding into the decision-making process of a fully connected layer.
9.3.4 Activations
Throughout the previous sections we have discussed features, convolutions, and dense layers, but have purposefully been neglecting a very important step in the process. Part of the true power in deep learning comes not only from these critical parts, but also from the nal step after each: the activation function.
9.3.4.1 Linear activations
The activation function is a deceivingly simple thing. A simple statement can summarize its purpose: using a dened function, map an input value to an output value. One of the most basic forms of activation is a linear activation: f(x) = mx + b, which can have a scaling (m) and bias (b) impact on the outputs. While linear activations can be a useful way of scaling and shifting values, stacking multiple linearly activated layers does not add any new information to the system. To demonstrate this, imagine two linear feed-forward layers, where x is the original values:
=+ymxb
11
1
=+ymyb.
2
2
2
1
The entirety of this network y2can be dened and simplied as
=++→+++ymmxbbmmxmbbMxB,
()
21 1 2 21 21 2
2
where mb1+ b2can be combined as a single bias, B, and mm1can be a single scalar, M. This means that regardless of how many linearly activated layers are placed together, they are only capable of expressing linear relationships. For modeling something with non-linear such as gravity, they would inevitably fall short.
9-8
Artificial Intelligence in Adaptive Radiation Therapy
9.3.4.2 Non-linear activations
The power of multiple layers becomes more apparent as we transition towards non­linear activation functions. Stacking multiple non-linearly activating layers can add increasing complexity to the models predictive abilities. For example, if the activation function were f(x) = mx representational ability of several polynomials: x
2
+ gx + c, two such layers would have the
4
, x3, x2, x, and bias.
A set of four commonly seen non-linear activation functions is shown in gure 9.7.
Figure 9.7. List of common activation functions, sigmoid, rectied linear unit (ReLU), leaky ReLU, and exponential linear unit (ELU).
9.3.4.3 Sigmoid activation
The sigmoid activation function, while not the most commonly seen in the intermediate layers of deep learning architectures, is a common nal activation function for a binary prediction. This function is commonly expressed as being synonymous with neurons ring within the brain. Given a number of inputs, if the value exceeds some critical threshold the neuron will re. A sigmoid activation is an appropriate activation for a binary classication system such as our dog model shown in gure 9.6, where 0would indicate no dog, and 1would indicate the presence of a dog.
In the case of multiple classication options, the sigmoid activation is often exchanged for the soft-max activation. This provides a probability of a class scaled to the sum of all classs probability:
z
c
==
ycx
∣)
(
C
=
j
e
,
z
j
e
1
where c is a particular class and C is the total number of classes.
9.3.4.4 Non-linear activations: rectified linear
The rectied linear unit activation (ReLU) was one of the rst and most popular activation functions used in modern deep learning architectures. The function is relatively simple: if x is less than 0, the returned value is 0, otherwise the returned value is x. The leakyReLU and exponential linear unit (ELU) are both variations
9-9
Artificial Intelligence in Adaptive Radiation Therapy
on the ReLU, where the differences focus on what occurs when values less than 0 are presented to the function. One of the major issues with ReLU was dyingnodes, meaning if a kernel initialized at a value less than 0, all gradient through the node died. The leaky and ELU activation allow at least some gradient.
An important question to raise at this point is Why use these non-linear activations?The answer to this lies in the most interesting part of deep learning: loss and back-propagation.
9.3.5 Loss: driving the model
When making a prediction, we need some method of identifying the correctnessof the model prediction. This metric is referred to as the loss. There are several ways of expressing a model loss, although, for mathematical reasons, it is desired that this loss is something that the model is attempting to minimize. An in-depth discussion of back-propagation and gradient descent is beyond the scope of this chapter, please refer to other texts discussed at the beginning of this chapter [25, 26] for a detailed explanation.
9.3.6 Auto-encoders: remove the noise
If one were to be told to draw a bicycle, remove while playing a game of Telestrations, they would likely draw out two simple circles as wheels, a frame connecting them, and a set of handlebars. Perhaps peddles and spokes on the wheels would be included for the more artistically inclined, and yet, this drawing of a bicycle would be relatively far removed from the appearance of an actual bicycle. While there are hundreds, if not thousands, of different types of bicycles, most people would be able to review this crude drawing and conclude that the image is supposed to be that of a bicycle. The human brain is amazing in this regard: a bicycle, in all its many shapes and forms, can be condensed down to just a few simple aspects. This condensing of important features or removing of unnecessary parts (noise) is the basis of auto-encoders [27].
The goal of an auto-encoder is to simplify, or compress, a set of inputs into its most important aspects, gure 9.8. By reducing the number of features and accurately reproducing the input, the model is able to identify what are the most important aspects of the input. This can similarly be used as a form of principal component analysis (PCA) [28] and is a common building block of both supervised and unsupervised learning models. The so-called stackedauto-encoder is simply a combination of auto-encoders with signicantly noted improvement in model predictive capability [29]. The training of these models is often pieced together, with each layer trained separately before being combined and ne-tuned in the nal model.
9.3.7 Supervised versus unsupervised learning
Distinctions between supervised and unsupervised learning can be simplied as a statement about data labeling. If the user understands what the desired output should be, e.g. is a stop sign present in this image?and uses those labels to train a
9-10
Artificial Intelligence in Adaptive Radiation Therapy
Figure 9.8. Basic example of an auto-encoder.
model, we are operating under supervisedlearning. However, if we simply had many pictures of stop signs and other signs with no ground truthlabeling, we would require an unsupervised learning technique. This is not to say that unsuper-
vised learning does not require an equal amount of work for training as a supervised learning model.
9.3.8 Pre-trained convolutional neural networks
Much of the previous discussion revolves around feature extraction (convolutions), renement (pooling), and combination of those features (dense layers), all with the understanding that relevant features are being identied by the model. However, this is not a guarantee. When rst creating a convolutional network, such as one for image classication by the visual geometry group 16 (VGG-16) [30], gure 9.9, every convolutional kernel is going to be a randomly generated distribution of numbers.
Deep learning researchers all suffer from this issue. Updating these random kernels based on correct predictions is a labor intensive and sensitive process that occupies most, if not all, of a researchers time and effort when beginning a new classication process. The VGG-16 model was trained on the ImageNet challenge, which contains more than 14 million images. This is far above what most medical researchers have available for training their own models. However, this can still be used to our advantage with the concept of using this pre-trained model.
9-11
Artificial Intelligence in Adaptive Radiation Therapy
Figure 9.9. Representation of the visual geometry group (VGG-16) classication architecture.
The fundamental basis of using a pre-trained model is that certain low-level features are ubiquitous. As humans, we are able to identify many different objects all of which are built on the same fundamental pieces: lines, curves, edges, etc. We are then able to take this pre-trained model that has been trained on thousands of different images and utilize everything before the fully connected layers as a feature extraction network. This saves potentially countless hours of training; by freezing the convolution layers, we can simply train the end classication. Note: Even after a new classication model is trained, it is common to un-freeze the earlier layers in a process of ne tuning.

9.4 Image registration: bringing two images together

Throughout the course of a patients care there is a potential for multiple forms of medical imaging to be acquired for various purposes, e.g. dening tumor bounda­ries, identifying surrounding critical normal tissue, and characterizing tumor and normal tissue function. For example, a magnetic resonance image (MRI) can exhibit exquisite contrast of soft tissues in ways that computed tomography (CT) scans cannot, and an FDG positron emission tomography (PET) scan is able to identify regions of metabolic hyperactivity. Currently, standard practice within radiation oncology includes a CT scan enabling the visualization of a desired treatment site and enabling radiation dose calculation, although there is signicant research into MRI based generation of electron density.
An ability to refer to multiple imaging modalities to evaluate tumor extent or normal tissues is vital to the decision-making and treatment planning process. Unfortunately, these PET, MRI, and CT scans are rarely acquired at the same time or with exactly consistent positioning of the patient. These changes in time, positioning, and/or anatomy have led to the necessity of tools to align different modalities/images about a region of interest in the anatomy. This aligning of different modalities or images is referred to as image registration.
This is not to say that image registration is only important during the planning process. Accurate alignment of the planning images with treatment imaging (cone­beam CT, fan-beam CT, MRI) is equally vital to a safe and successful treatment.
9-12
Artificial Intelligence in Adaptive Radiation Therapy
With ART, it is often vital to understand the distribution of previously delivered radiation for guidance in future planning. This understanding relies on accurate representation and registration of the new imaging with previous plans. It is equally important to understand how a registration is evaluated and functions, the registration metric, per AAPM Task Group 132 [31]. After all, when registering two images, how does one know when the registration is good?
There are multiple ways in which a registration can be evaluated. Often, the registration is focused on a particular region of clinical relevance. Identifying the region of relevance is the rst and most important aspect of image registration.
9.4.1 Registration similarity metrics
There are broadly two ways of evaluating/driving a registration: intensity-based and feature-based.
9.4.1.1 Intensity based registration
Intensity based registration can be dened as a comparison of intensities between the two images, either as a whole, or about a dened sub-region of interest in the image [3234]. When operating between two images of the same modality, simple comparisons of the intensity values between the two can offer a reasonable evaluation of the registration, possible as the sum of squared differences (SSD) [35, 36]:
N
1
=−
()
IISSD
XY
N
=
i
1
2
ii
,
where I is the intensity and subscript X and Y refer to two images across the total number of voxels being evaluated. While this metric can be very useful, it suffers when the voxel intensities throughout the two images vary (a multi-phase contrast enhanced CT). To this end, the normalized cross-correlation (NCC) [3638]isa wonderful metric, accounting for the differences in intensity between the two images, although assuming that relatively high and low intensities correlate with each other:
n
−−
xxyy
()()
i
=
i
=
CC .
1
n
−−
xx yy
()()
i
∑∑
==
i
1
i
n
2
i
1
2
i
An important note from both methods is that they require the intensities between the images to be correlated: a bright spot in one image corresponds to a bright spot in the other, etc. For multi-modality images (CT to MR), this is not a guarantee.
A commonly used metric for the registration of images of different modalities is mutual information [39, 40]. This metric is based on the mutual probabilities between the two images and has no reliance on the absolute intensity of the images:
9-13