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
IOP Publishing
Artificial Intelligence in Adaptive Radiation Therapy
Yi Wang and X. Sharon Qi
Chapter 1
Fundamentals of artificial intelligence
Parsa Bagherzadeh, Laya Rafiee Sevyeri, Yujing Zou and Shirin Abbasinejad Enger
Articial Intelligence (AI) is transforming healthcare, with adaptive radiotherapy standing out as a key area where AI is reshaping cancer treatment. By leveraging real-time data, AI enables more personalized and precise treatment plans, improving patient outcomes. This chapter introduces the fundamentals of AI, focusing on machine learning (ML) and deep learning (DL), which drive innovations in radio­therapy. It covers the evolution of AI from rule-based systems to advanced learning algorithms, exploring their applications in clinical settings. It will discuss the ethical considerations surrounding AI in healthcare, including transparency, bias, and data privacy, setting the stage for a deeper dive into AIs role in adaptive radiotherapy.

1.1 A brief introduction to AI

Articial intelligence (AI), a subeld of computer science, focuses on developing systems capable of performing tasks that typically require human intelligence. Such tasks include problem-solving, learning, perception, and language understanding. AI systems leverage algorithms, statistical models, and machine learning techniques to analyse and interpret data, enabling them to make decisions, recognize patterns, and improve performance over time. AI applications range from virtual assistants and recommendation systems to complex tasks in various elds. In particular, AI plays a signicant role in healthcare, transforming medical decision-making and treatment strategies by changing the existing landscape. In the context of adaptive radiotherapy, an intricate interplay of AI methodologies is reshaping how we approach cancer treatments. The term articial intelligencewas coined at the Dartmouth Conference in 1956, where researchers aimed to explore how machines could simulate human intelligence. In the 1960s and 1970s, AI research mainly focused on logic-based symbolic reasoning, which involves representing knowledge using symbols and rules, allowing machines to manipulate these symbols to derive logical conclusions and make decisions [1]. During this period, AI found its
doi:10.1088/978-0-7503-6119-4ch1 1-1 ª IOP Publishing Ltd 2025. All rights,
including for text and data mining (TDM), artificial intelligence (AI) training, and similar technologies, are reserved.
Artificial Intelligence in Adaptive Radiation Therapy
applications in medicine through systems such as MYCIN [2], a system for diagnosing bacterial infections and recommending antibiotic treatments.
In parallel with symbolic reasoning, AI research also focused on optimization and handling uncertainty, both important for developing intelligent systems. Fuzzy logic, for instance, is a mathematical framework that deals with uncertainty as well as imprecision, allowing for the representation of partial truths [3]. This exibility makes it particularly suitable for systems where variables may have ambiguous or overlapping boundaries. Genetic algorithms are examples of evolutionary optimi­zation techniques inspired by principles of natural selection and genetics [4]. They employ a population of potential solutions, subjecting them to genetic operations such as crossover, mutation, and selection to evolve toward an optimal solution over successive generations.
The development of reasoning systems continued in the 1980s as expert systems. Expert systems serve as digital counterparts to human domain experts, leveraging a comprehensive knowledge base to guide decision-making processes. These systems bring a wealth of rules and domain-specic knowledge into play, enhancing the precision of decision-making [5].
ONCOCIN [6] is an example of a system designed to assist oncologists in the treatment of cancer patients. Note that the rules in expert systems are predened and programmed into the system, outlining how the system needs to behave in different situations. Expert systems, thus, lack the ability to adapt or learn from new data, and they operate within the boundaries of the predetermined rules set by human experts and the initial knowledge programmed into the system.
The limitations of expert systems later motivated the development of machine learning (ML) models in the 1990s and 2000s. In contrast to expert systems, ML models extrapolate from data. ML involves the development of algorithms that enable machines to learn and improve from experience. In adaptive radiotherapy, ML algorithms can leverage extensive patient data to discern patterns and forecast customized treatment plans [7]. For example, these algorithms can be used in segmentation tasks to help delineate volumes concerning the tumor and organs at risk. This progresses to treatment management, addressing motion during the treatment and treatment plan optimization. Ultimately, decisions are interconnected through reviewing treatment specics, enabling adaptive changes.
While classical ML models do not have predened rules such as expert systems, they still require human expert efforts to design a feature set. These features, specic attributes or characteristics selected from input data, provide structured information to help algorithms discern and differentiate between data inputs, enhancing pattern recognition and prediction capabilities. Deep learning (DL), a subeld of ML, employs neural networks with multiple layers to extract representations from raw data [8] without explicit feature engineering. DL has dominated AI research since
2010. DL, with its capacity to autonomously acquire hierarchical features from intricate data, has the potential to enhance precision and efciency across all stages of radiotherapy treatment and its automation.
The landscape of AI in healthcare has undergone a transformative shift with the rise of ML and DL techniques. This chapter focuses on ML and DL, describing their
1-2
D
X
y
y
D
Artificial Intelligence in Adaptive Radiation Therapy
capacity to discern intricate patterns, adapt to dynamic patient conditions, and optimize and automate radiotherapy treatments. The nuances of these data-driven approaches within the context of adaptive radiotherapy are explored. A practical understanding of these AI approaches is provided from classical ML to DL and generative models. An overview of ethical considerations regarding transparency and bias is also provided.

1.2 Machine learning basics

This section briey introduces basic denitions of ML, including different learning strategies, feature representation, feature selection, and classical ML models.
1.2.1 Learning paradigms
An ML algorithm refers to a type of algorithm that can learn from experience. The learning process can occur in various ways. Considering the data and the type of information it provides, three main learning paradigms can be dened: (i) super­vised, (ii) unsupervised, and (iii) self-supervised learning.
Supervised learning is the most commonly used ML paradigm. It earns its name as it resembles a teacher guiding a students progress. Here, the algorithm learns from a training dataset, with known correct answers, by making predictions iteratively, which are then corrected. This process continues until the algorithm achieves a satisfactory performance level. Formally, supervised learning involves training a model on a dataset (
) is paired with its corresponding ground-truth label (yi∈ Y). In summary,
x
i
=…
xy xy xy,,,,,,
train 1
{( ) ( ) ( )}
2
1
2
the goal of supervised learning is to learn a function f that accurately maps input an output
, where
=
provides a reliable prediction ofy[9].
fx()
Unsupervised learning assumes that access is limited to the input variables, and their corresponding ground truth remains unavailable ( objective of unsupervised learning is to model the underlying structure or distribu­tion of the data to gain insights. Unlike supervised learning, no supervision is provided for the given training data. Unsupervised learning problems extend beyond traditional clustering, which involves grouping similar samples into distinct clusters, and association problems, which discover interesting relationships or patterns among variables in large datasets. Unsupervised models include dimensionality reduction models, such as principal component analysis (PCA) [10], as well as generative models such as autoencoders [11] and generative adversarial networks (GANs) [12], which are widely used to generate examples resembling the training data.
Self-supervised learning is a fairly new ML paradigm where a model learns from the data without relying on externally provided labeled annotations. Instead of using labeled data created by human annotators, self-supervised learning leverages the inherent structure or information within the data to generate supervisory signals. The model is tasked with predicting certain parts or properties of the input data based on other parts, creating a pretext task such as jigsaw puzzle [13] or rotations detection [14]. This process enables the model to learn useful representations or
, where each sample
n
n
=…
xx x,,,
{}
). The
ntrain 1 2
x
to
1-3
Artificial Intelligence in Adaptive Radiation Therapy
features from the input data, which can later be utilized for downstream tasks such as classication or regression. Self-supervised learning is particularly valuable in scenarios where obtaining labeled data is challenging or expensive.
In addition to these three paradigms, one may also consider semi-supervised learning and reinforcement learning paradigms. In semi-supervised learning, we typically have access to a small set of labeled data along with a large set of unlabeled data, while in reinforcement learning an agent learns a task by interacting with an environment and updates its approach based on the positive or negative rewards it receives.
1.2.2 Regression versus classication
Classication and regression are two fundamental types of supervised ML tasks. Classication aims to assign input data points to predened categories or classes, as illustrated in gure 1.1(a). Classifying tissue samples as normal, benign, or malignant based on histopathological images, identifying the presence or absence of specic genetic mutations, distinguishing between different stages of cancer progression based on imaging data, and detecting cancer recurrence or metastasis from medical imaging scans, are examples of classication tasks in precision oncology. In contrast, in regression the goal is to predict a continuous numeric output based on input features, as shown in gure 1.1(b). A regression model may use features such as patient age, sex, tumor size, Gleason score, stage, and grade to predict radiation dose, which is a continuous numeric value.
1.2.3 Feature engineering and representation
Representing real-world samples using structures understandable by computers (often vectors or matrices) is the rst step in developing an ML model. Considering the task at hand, domain experts dene a set of features that capture the most important characteristics of a sample. Such features serve as the descriptors that constitute the input for ML algorithms. This step in developing ML models is often referred to as feature engineering and representation. A sample dataset for the
Figure 1.1. Classication versus regression. (a) A sample classication example of two classes with a clear decision boundary. (b) A regression model tted on a sample dataset.
1-4
Artificial Intelligence in Adaptive Radiation Therapy
Table 1.1. A sample dataset for predicting the therapeutic role of radiotherapy.
Time since
ID Age Sex
1 45 Male 120 2.5 Chest Lung T2 Curative 2 62 Female 90 1.8 Abdomen Kidney T1 Palliative 3 55 Male 150 1.9 Neck Thyroid T3 Curative 4 70 Female 180 2.2 Pelvis Prostate T2 Palliative 5 38 Male 75 1.6 Extremities Arm T1 Curative 6 48 Female 110 2.3 Chest Breast T2 Curative 7 64 Male 200 2.8 Neck Tongue T3 Palliative 8 58 Male 130 2.1 Abdomen Pancreas T2 Palliative 9 50 Female 160 2.4 Pelvis Ovary T3 Curative 10 42 Male 85 1.9 Extremities Leg T1 Curative
1st RT
Dose/fraction (Gy) Region Site
T stage Output
task of predicting the therapeutic role of radiotherapy is provided in table 1.1. The expert has designed/engineered a set of d
= 8 features (the rst eight columns) as
in
predictive features for the output (the last column). Each entity in the world, in this case a patient, is then characterized using the feature set.
1.2.3.1 Feature encoding
As table 1.1 shows, the features are inherently different in terms of the values they can take. Features such as sex, region, site, etc, are nominal categorical and comprise a nite set of discrete values with no inherent ranking or order. An ordinal variable also comprises a nite set of discrete values, but it has a clear ranked ordering between the values, such as ratings (e.g. low, medium, high) or stages (e.g. T1, T2, T3). Another feature type is numerical features such as age, dose, time, etc, that take continuous values from a domain. As the name suggests, such features are characterized by numeric values and can represent a wide range of data types, such as integers, real numbers, or ratios.
Identifying and understating different feature types is important since it deter­mines which type of ML models might be applicable. Different ML models require different encoding schema for their inputs. As introduced in section 1.2.5, decision trees, for instance, assume that all features are categorical, while models such as k­nearest neighbors and support vector machines that operate on vector spaces require the features to be numerical.
Ordinal encoding, for instance, is used when the categorical values have a clear order or rank. It assigns a unique numerical value to each category based on its position in the order as shown in the feature stage in table 1.2. On the other hand, one-hot encoding is used for categorical variables that do not have a natural rank ordering. Assuming m different categories, each category can be encoded by a one­hot vector {0, 1}
m
, where each dimension corresponds to a category. For instance, in
table 1.2, the category Malefor sex is encoded as 1, 0.
1-5
Artificial Intelligence in Adaptive Radiation Therapy
Table 1.2. Different encoding schema.
Feature Categorical Numerical
Stage T1, T2, T3 1, 2, 3 Sex Male, female Age
R [1, 20), [20, 40), [40, 60)
>< >1, 0 , 0, 1
Converting numerical features to categorical features, on the other hand, often involves discretization of the domain into intervals, each of which is considered a category (see Agein table 1.2). When discretizing numerical features, an important design decision is the number of intervals, which can affect the models performance. For instance, if the number of intervals is too low, the categorical feature may oversimplify the underlying numerical distribution, losing important information and patterns. Discretizing numerical features can also be used for data de­identication [15]. This reduction in granularity serves as a privacy-enhancing mechanism, especially in medical applications where the disclosure of ne-grained details could lead to the re-identication of individuals.
1.2.3.2 Feature selection
Medical data are inherently complex, often characterized by many features, and patient information extends far beyond basic demographics, comprising variables such as genetic markers, vital signs, medical history, diagnostic results, etc. The number of features in adaptive radiotherapy is even larger since it requires monitoring changes over time. Patientsanatomy, for instance, may evolve during the course of treatment due to weight loss, tumor regression/progression, and each time point adds new features. Moreover, radiotherapy involves a large set of dosimetric features, including dose distributions and dose–volume histograms, which leads to an increase in the number of features.
Many features, however, can lead to computational inefciencies, resulting in long training and prediction times. Moreover, it might increase the risk of over­tting (discussed in section 1.4), as the model may capture noise and intricacies specic to the training data, hindering its ability to generalize to new instances. The complexity imposed by numerous features impedes computational efciency and hampers model interpretability. Thus, pre-processing is required to select a subset of features to mitigate these challenges and enhance the overall performance of the ML model. Feature selection involves choosing a subset of the original features from the dataset based on their relevance to the task at hand [16]. The goal is to retain the most informative features while discarding irrelevant or redundant ones, reducing the complexity of the model, and potentially improving its interpretability. As an example, in some datasets, features such as patient ID often carry random values without meaningful patterns for predictive purposes. Including such features may introduce noise and deteriorate the predictive ability. Therefore, careful feature selection is crucial to exclude random-valued features. Note that patient IDs may
1-6
)
Artificial Intelligence in Adaptive Radiation Therapy
contain some information. For instance, some databases might incrementally assign IDs, meaning that the ID is correlated with the date of diagnosis. In other cases where certain features are constant or nearly constant across a patient cohort, including all instances of these features, they may not add value and could be considered redundant. For instance, body site might be highly correlated with sex. An example is the ovary, specic to the female reproductive system, rendering the sex feature redundant. Finally, some features may bear some information for output prediction. Their contribution as strong predictors needs to be evaluated to avoid overemphasizing less impactful variables. Feature selection involves using various measures to assess the relevance and importance of features in a dataset. Common techniques include evaluating features based on feature correlation coefcients or statistical measures such as mutual information.
Pearson correlation, for instance, is a statistical measure that quanties the strength and direction of a linear relationship between two variables x and y:
n
−−
xxyy
()()
i
i
1
=
=
r
xy
n
−−
xx xy
()()
i
∑∑
i
1
==
i
n
2
i
i
1
,
2
()
1.1
wherexandyare the mean values forxandy. The correlation coefcient is a dimensionless value that ranges from 1 to 1. A coefcient of 1 indicates a perfect positive linear relationship, meaning that as one variable increases, the other also increases proportionally. Conversely, a coefcient of 1 signies a perfect negative linear relationship, where one variable decreases as the other increases. A coefcient of 0 implies no linear correlation between the variables.
Mutual information, on the other hand, can capture non-linear relationships between features. Assuming
,,
as possible values fory, mutual information
v
y1
∣∣
as possible values for featurexand
u,,
x1 ∣∣
is an information-
xy,(
theoretic measure which is dened based on individual and joint entropies of two random variables:
=+−IxyHxHyHxy,,
() () () ()
x
== =
Hx Px u P x u
() () ()
=
j
1
y
== =
Hy Py P y
() () ()
=
k
1
log log ,
jj
log log .
vv
kk
1.2
()
This formulation of the mutual information is only applicable to categorical features. For numerical continuous features, an estimate is used, which is imple­mented in most of the existing ML libraries such as scikit-learn [17].
Feature selection is a subset of the broader dimensionality reduction paradigm [18]. In fact, any feature subset selection is a dimensionality reduction but not vice versa. Some dimensionality reduction methods, such as PCA and autoencoders,
1-7
Artificial Intelligence in Adaptive Radiation Therapy
Figure 1.2. Two different geometric distributions of data: (a) linear separable and (b) linear non-separable.
transform the original feature space into a lower-dimensional representation by creating new composite features. In medical applications, feature selection might be preferable to such approaches to enhance interpretability. Feature selection retains a subset of the original features, providing transparency in the models decision­making process. This transparency is important in medical contexts where under­standing the relationship between input features and predictions is essential for gaining trust from clinicians and ensuring the responsible adoption of the model in clinical practice.
1.2.4 Linear separability
When designing an ML model, an important consideration is the geometric distribution of data in terms of linear (non-)separability. Linear separability and linear non-separability refer to the inherent structure of a dataset in relation to a classication task. In a scenario of linear separability, classes within the data can be effectively separated by a hyperplanea geometric construct in the feature space as presented in gure 1.2(a). This means that a straight line or plane can distinguish between different classes, allowing the application of linear classication algorithms such as linear support vector machines or logistic regression. On the other hand, in cases of linear non-separability, a single hyperplane cannot accurately separate the classes, requiring more complex decision boundaries, such as curves or non-linear surfaces as illustrated in gure 1.2(b). In such cases, non-linear classication techniques such as kernel methods or DL models may be more appropriate to capture the intricate relationships within the data. Understanding whether a dataset exhibits linear separability or not is important in selecting the most suitable ML approach for a given classication problem.
1.2.5 Classical models
Classical ML models have been foundational in the evolution of AI, serving as powerful tools for data analysis and decision-making across various domains,
1-8
Artificial Intelligence in Adaptive Radiation Therapy
including healthcare. These models rely on predened features as described in section 1.2.3 and statistical techniques to identify patterns and make predictions. Understanding these models is important since it provides a knowledge of the principles and methods widely used in the eld, enabling practitioners to appreciate the evolution of ML.
1.2.5.1 Decision trees
Decision trees are a fundamental concept in ML, widely employed for their intuitive and transparent representation of decision-making processes. Serving as predictive models, decision trees navigate through a series of (often binary) choices based on input features to reach a nal decision or prediction. Figure 1.3 shows a decision tree learned based on the dataset provided in table 1.1. Decision trees are composed of three types of nodes: root, internal nodes, and leaves. All root and internal nodes correspond to features, and depending on the feature value, one of the sub-trees is traversed. The sub-trees are iteratively traversed until a leaf node is reached, corresponding to a classication decision.
The training process of a decision tree is often referred to as tree induction [20]. During the induction, at each root/internal node, the decision tree tries to partition the training samples (based on feature value) such that the subsets have the least impurity regarding class labels. The class impurity can be quantied using measures such as entropy or Gini index. Assuming a binary classication problem, the maximum entropy and Gini values are 1.0 and 0.5, respectively. The minimum value for these measures often occurs for the leaf nodesthe least impurity and the most certainty for the classication. In some cases, however, the designer might prefer to
Figure 1.3. A decision tree for predicting the therapeutic role of radiotherapy (inspired by [19]). Based on the value of features, the branches are iteratively traversed until a leaf node is reached.
1-9
Artificial Intelligence in Adaptive Radiation Therapy
terminate the induction process to keep the tree below a certain height1(for instance, to avoid over-tting). In such cases, some leaf nodes might have a non-zero impurity value. Note that the induction process may ignore some of the features. For instance, in gure 1.3, features such as sex or age do not appear in the tree. This means that the decision tree has implicitly performed a feature subset selection.
1.2.5.2 Random forests
Decision trees, while powerful, suffer from limitations. Their tendency to overt (discussed in section 1.4.4) and sensitivity to data variations can lead to unreliable predictions. Moreover, decision trees have a greedy optimization scheme and they have limited ability to capture complex relationships.
Random forests represent an ensemble machine learning method that leverages the collective power of multiple decision trees. Each tree is constructed from a random subset of features, introducing an element of stochasticity that enhances generalization. By aggregating the predictions from this ensemble, random forests achieve robustness to over-tting and improve overall prediction accuracy compared to individual decision trees. This technique has gained widespread adoption due to its effectiveness in various classication and regression tasks.
1.2.5.3 Support vector machines
Assuming a linear separable distribution for data, the classes can be separated
d
using a line or hyperplan e characterized as
+=
wb0.
,where
in
wR
and b
R
are the slope and intercept parameters of the hyperplane. There are, however, innite possible values for the parameters, resulting in many potential discrim­inators as shown in gure 1.4(a). While all t he hyperplanes effectively separate the training samples, some of them may misclassify test samples, leading to poor generalization.
Figure 1.4. Support vector machine. (a) Several possible separating lines and (b) the separating line with the largest margin.
1
The height of a tree is dened as the maximum number of internal nodes on the longest path from the root to
any leaf node.
1-10