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
performance on new, unseen data. Thus, in addition to a training set, two other hold-out sets, namely validation and test sets, are used. While a training set directly inuences the learning process of a model, a validation set is used to evaluate the performance of the model during the learning, to tune hyperparameters, and select the best performing model. The test set on the other hand is never disclosed to the model during the learning process and is only used to evaluate the performance once the training is nished. In ML it is common to have a small sized dataset. This might pose a challenge when splitting the data, since the performance estimates might be misleading due the small size of partitions. Techniques such as k-fold cross-validation and stratied k-fold cross-validation mitigate issues related to small or moderate-sized datasets, offering more accurate estimates of general­ization error [39].
1.4.3 Evaluation metrics
Once an ML model is developed, its performance needs to be evaluated, using quantitative measures. The performance of a predictive model is often measured on a hold-out test set, i.e. a set that has not been exposed to the model during its training process. This allows informed decisions to be made in comparing different models and choosing the one that has the best performance for unseen data (generalization capability).
The confusion matrix provides detailed insights into predictive model perform­ance, which shows correct and incorrect predictions for each class. Figure 1.12(a) shows a sample confusion matrix for a binary classication problem. Each entry in the matrix counts the number of correct/incorrect classications for each class. True positive (TP) is the number of positive test samples correctly classied as positive.
Figure 1.12. Performance evaluation: (a) a confusion matrix and (b) an ROC curve.
1-21
Artificial Intelligence in Adaptive Radiation Therapy
Likewise, true negative (TN) is the number of negative samples correctly classied as negative. False positive (FP) corresponds to the number of negative test samples misclassied as positive, and false negative (FN) is the number of positive samples misclassied as negative.
A set of performance measures can be dened to aggregate the matrix entries and provide more focused insight into a models performance, especially in cases of imbalanced distributions. Three of the most common metrics are recall, precision, and F1 score.
Recall is the ratio of true positive predictions to the total number of actual positives:
TP
=+R
TP FN
.
1.9()
Recall measures the ability of a model to capture all positive instances. High recall means that the model has fewer false negatives. In the example provided in gure 1.13(a) the model has few false negatives (FN = 3) leading to a relatively high recall (R = 0.84). Note that a model can have a perfect recall (R = 1) by predicting all samples as positive (i.e. FN = 0).
Precision, on the other, is the ratio of true positive predictions to the total number of positive predictions:
TP
=+P
TP FP
.
1.10()
Precision measures the accuracy of positive predictions, and a high precision means that the model has fewer false positives. In gure 1.13(a), while the model demonstrated high r ecall, it shows a poor precisi on (only slightly better than random) due to the relatively large number of f alse positives. The emphasis on precision or recall depends on the domain and the task at hand. F or instance, when identifying treatment options, it is crucial to have a h igh precision while for identifying hi gh-risk patients for screening, a high recall is important.
Figure 1.13. Different tting scenarios: (a) over-tting, (b) proper tting, and (c) under-tting.
1-22
Artificial Intelligence in Adaptive Radiation Therapy
The F1 score combines precision and recall using their harmonic mean:
×+PR
=
F1 2 .
PR
The F1 score provides a balanced measure of a models performance by considering both false positives and false negatives. It is particularly useful in cases with an imbalanced class distribution or when both types of classication errors have signicant implications. Figure 1.13(a) shows that the model has a moderate F1 due to low precision.
The receiver operating characteristics (ROC) curve is a valuable tool for assessing a models ability to discriminate between classes, visualizing the tradeoff between the false positive rate (FPR) and true negative rate (TPR) (see gure 1.13(b)). The area under the curve (AUC) quanties model performance, with perfect prediction yielding an AUC of 1 and random prediction an AUC of 0.5. As gure 1.13(b) shows, the AUC for classier 1 is bigger than the AUC for classier 2, suggesting a better performance for classier 1.
1.4.4 Overtting versus under-tting
The development of an ML model involves striking a balance between complexity and simplicity. Neglecting this balance may lead to under-tting or over-tting. Under-tting means the model is too simplistic to capture the underlying patterns in the training data, e.g. using a linear classier for linear non-separable data (see gure 1.13(a)). An indicator of under-tting is a low performance or high error rate on the training data, let alone test data. This is because the model fails to grasp even the most basic patterns present in the training data.
Conversely, over-tting occurs when the model has a high learning capacity, and the training data has a relatively simple distribution (see gure 1.13(b)). High performance (low error) on training data, but a low performance on test data are indicators of over-tting. Hyperparameter tuning is a common approach to prevent over-tting. In decision trees, for instance, techniques such as pruning, which involves limiting the growth of the tree by setting a maximum depth or restricting the minimum number of samples required to split a node can be used [20].
1.11()

1.5 Generative models

Generative models are a class of ML models designed to learn and mimic the underlying distribution of a given dataset. Unlike discriminative models, which focus on predicting labels or classifying data (p(yx)), generative models aim to capture the joint probability distribution of the input data and the corresponding labels (p(x, y)).
The primary purpose of generative models is to generate new data samples that are similar to the training data. These models can generate new examples from scratch by sampling from the learned distribution, allowing them to create realistic data that preserves the statistical properties of the original dataset. Generative models have various applications, including data augmentation, image and text
1-23
Artificial Intelligence in Adaptive Radiation Therapy
synthesis, anomaly detection, and semi-supervised learning. They are particularly useful in scenarios where obtaining labeled data is expensive or impractical, as they can generate synthetic data for training discriminative models. While various deep generative models have been proposed in the literature, we focus solely on introducing the two most prominent and recent techniques in this discussion.
1.5.1 Generative adversarial networks
A generative adversarial network (GAN) [12] is a deep generative model where its primary goal is to mimic the distribution of training data and, consequently, generate samples drawn from the learned distribution. GAN is a two-player minimax game involving two opposing models: a generator G and a discriminator D, as shown in gure 1.14. In this framework, both models are trained simulta- neously. The discriminator aims to distinguish between samples from the true data distribution and the generator distribution. In contrast, the generator seeks to minimize the likelihood of being identied as fake by approximating the data distribution from a simpler distribution, such as Gaussian or uniform.
During training, the generator tries to mislead the discriminator, while the discriminator endeavors to maximize the probability of accurately predicting true labels for both real and generated samples. The competitive dynamic between these two components encourages continual improvement. The ideal stopping point is reached when G captures the distribution of the training data (p
g
= p
), and D can
data
no longer differentiate between generated and training data, resulting in a proba­bility close to 0.5 for each sample.
Given G and D as neural networks, the training of GANs is formulated as follows. The generator G aims to learn the distribution over data x and a prior on the input noise variable, dened as p neural network parameterized by θ function for the discriminator D. Here, p
(z). G(z; θg) is the mapping function, where G is a
z
. Conversely, D(x; θd)defines the mapping
g
and p
g
represent the generated and
data
training data distributions, respectively. The discriminators output is a scalar representing the label of the input data; that is, D(x) indicates the probability of x belonging to the training data distribution rather than p
. While D maximizes the
g
probability of assigning true labels to both training and generated samples, G minimizes log(1 D(G(z))). Thus, D and G engage in a minimax game to optimize the following objective function:
Figure 1.14. GAN architecture; z represents noise sampled from a Gaussian distribution, G(z) denotes the generated image from the noise z, and x represents a training sample drawn from the p (Reproduced with permission from [
42].)
distribution.
data
1-24
)
0
Artificial Intelligence in Adaptive Radiation Therapy
∼= +
VDG E Dx E DGz, log log log log 1 .
( ) [ ( )] [ ( ( ( )))]
∼∼
xP x zPzz
() ()
data
1.12
()
Given that discrimination is inherently easier than generation, this objective function might result in a suboptimal generator. Goodfellow et al [12] proposed reformulat­ing equation (1.12) by replacing the minimization of log(1 D(G(z))) with the maximization of log(D(G(z))).
This zero-sum game between these two components can lead to rich representa­tions of the training data, which can be further utilized for downstream tasks [40,
41]. While GANs have demonstrated great success in generating high-quality
samples, they have limitations. Due to their adversarial training nature, they are known for potentially unstable training and limited diversity in generation.
1.5.2 Diffusion models
While existing deep generative models excel at image generation, they encounter certain challenges. A diffusion model [43], belonging to the class of generative models, was introduced to address these challenges by generating high-delity images. Diffusion models contain two main processes: forward diffusion process and reverse diffusion process. They dene a chain of diffusion steps to slowly add random noise to data and then learn to reverse the diffusion process to construct desired data samples from the noise.
In the forward diffusion process shown in gure 1.15, we slowly and gradually add Gaussian noise to the input image x sampling a data point x
from the real data distribution q(x)(x0∼ q(x)) and then
0
add some Gaussian noise with variance β with distribution
(
−xt 1
through a series of T steps. We start with
0
to x
t
, producing a new latent variable x
t1
t
1
bb=−
t1: 0
1
qx x Nx x I;1 ,
(∣ ) ( )
−−
tt t tt t11
T
qx x qx ,1.13
(∣) () ()
T
where The data sample
Figure 1.15. Diffusion process: through the forward process, noise is added to a given image x, drawn from q (x), through T steps. respectively. Through the reverse process, the given image x is reconstructed through a denoising process from step T to 0.
andNdenote the identity matrix and Gaussian distribution, respectively.
gradually loses its distinguishable features as the step increases.
x
0
and
x
represent the original image and the image after T steps of adding noise,
x
T
=
=
t
1-25
p
)
Artificial Intelligence in Adaptive Radiation Therapy
Eventually, when
, it becomes equivalent to an isotropic Gaussian
→∞T
distribution.
The reverse diffusion process involves training a neural network to restore the original data by reversing the noise applied during the forward pass (see gure 1.15). Estimating q(x
) can be challenging as it can require the entire dataset. To
t−1xt
address this, the reparameterization technique is employed, utilizing a parameterized model p small β
in the form of a neural network to learn the parameters. For sufciently
θ
, the distribution becomes Gaussian, allowing the mean and variance to be
t
parameterized easily:
T
=
xpxpxx
() () ( ∣)
TT
qq
px Nx xt xt;,, ,.
() ( ()())
ttt t1
qq
=
t
m
tt0:
1
1
1.14
q
()
The network is trained to predict the mean and variance for each time step. Here
xt,t()
q
and
Σqxt,t(
dene the mean and the covariance matrix, respectively.
There are different varieties of diffusion models such as diffusion probabilistic models (DPM) [45], noise-conditioned score network [44], and denoising diffusion probabilistic models (DDPM) [45].
1.5.3 Applications and use cases
Generative models such as GANs, autoencoders, variational autoencoders, and diffusion models have been used extensively in the literature for the medical domain, mainly for image generation to mitigate the limited and imbalanced data problems in developing AI tools for medical problems [42]. Han et al [46] utilized the DCGAN and Wasserstein GAN (WGAN) [47] for medical image generation. Nie et al [48] suggested an adversarial model to generate magnetic resonance imaging (MRI) images from computed tomography (CT) images, whereas recently the authors of [49] suggested using denoising diffusion models to generate high-quality MRI and CT. GANs have also been used for synthetic generation in structured data [50].
Aside from image generation, transfer learning and domain adaptation models using generative models such as GANs have found widespread application in the medical eld. Domain adaptation is a subcategory of transfer learning when we aim to learn a model from a source data distribution and apply that model to a different target data distribution. Yaqoub et al [51] proposed using transfer learning in MR image reconstruction trained with GANs to mitigate the limited data problem in medical imaging tasks. Recent studies have applied domain adaptation approaches to address distribution shifts while leveraging existing medical data. Several studies considered working on MR images as the source domain and CT images as the target domain using adversarial training to generate synthetic CT images [52], cardiac structure segmentation [53], and image registration [54].
Generative models have also been used for anomaly detection. Anomaly detection, also known as out-of-distribution detection, focuses on recognizing samples that diverge from the rest of the data, signaling variations in measurement,
1-26
Artificial Intelligence in Adaptive Radiation Therapy
experimental errors, or novel occurrences. This proves valuable in uncovering unknown anomalies in the medical eld, where acquiring a suitable annotated dataset is consistently a challenge. This approach is also relevant in situations where information about the specic types of anomalies is scarce. The rst anomaly detection model in medical imaging using GAN was proposed to nd anomalies in optical coherence tomography [55]. While GANs were used mostly for detecting anomalies a few years ago, more recently, diffusion models have become the dominant approach. Wolleb et al [56] proposed a novel weakly supervised anomaly detection method based on denoising diffusion models to address the difculties of anomaly detection models based on GANs in preserving ne details in the image. Pinaya et al [57] proposed a method based on diffusion models to detect and segment anomalies in brain imaging. Aside from images, anomaly detection using generative models in time-series data also gained attention [58].
In addition to potential applications of generative models, they gained attention in other medical imaging tasks. Wolterink et al [59] proposed an adversarial model by training two separate generators with two different losses and combining them to reduce the noise in low-dose CT images. Image super-resolution using GANs has also been investigated in medical images [60]. Different generative models have also been used extensively for image and video segmentation. Wu et al [61] combined diffusion and transformer models for medical image segmentation. In another study, a denoising diffusion probabilistic model was integrated into standard U-Net [62] models for medical image segmentation [63]. Generative models, specically GANs, have been used for other modalities such as text. Even though generating discrete data such as text with GANs is challenging, several studies investigated using them for text generation purposes [64].

1.6 Ethical consideration and bias

AI has signicantly transformed various aspects of our lives, holding great potential for societal benets, ranging from healthcare applications to biomedical data analysis. In healthcare, AI has the capacity to enhance diagnostic accuracy, personalize treatment plans, improve patient outcomes, and streamline administra­tive tasks. Despite these positive impacts, there is also the potential for unintended harm and misuse. In the following, some of the most important of these potential harms and misuses will be explained.
1.6.1 Transparency and explainability
A computational system is considered transparent when every detail of its operation is known, whereas we refer to a system as explainable when humans can understand how it makes decisions. Transparency in a system does not necessarily ensure explainability. Understanding how a model operates can be particularly challenging, especially in the case of DL models with many parameters.
While explainability in DL models is often achieved through techniques such as layer-wise relevance propagation or attention mechanisms [30] to enhance their interpretability or model-agnostic approaches such as LIME [65], many classical
1-27
Artificial Intelligence in Adaptive Radiation Therapy
ML models, such as decision trees, SVMs, rule-based systems, and symbolic AI, inherently offer a certain level of explainability
1.6.2 Bias and fairness
Bias typically denotes a statistical deviation from a dened norm. In AI applica­tions, this divergence often arises from illegitimate or unrelated factors inuencing the output. For instance, gender is irrelevant to job performance; therefore, using gender as a basis for hiring a candidate is considered irrelevant (example: Amazon recruiting ML tool)
3
. Similarly, race is unrelated to criminality, making it irrelevant to incorporate race as a feature for predicting recidivism [66]. While bias can take various forms in AI models, it is mainly inuenced by our data selection strategy.
Data selection bias occurs when the training data are incomplete and comprise only specic distributions, such as race, gender, ethnicity, etc, making algorithmic bias likely. In a study by Obermeyer et al [67], an algorithm used in healthcare to identify and assist patients with complex medical needs was investigated. The study revealed racial bias in the algorithm, as it systematically underestimated the healthcare needs of black patients compared to white patients. In another study, commercial gender classication systems were evaluated [68]. The ndings showed that these systems exhibited higher accuracy for lighter-skinned and male faces compared to darker-skinned and female faces. The reason behind both models biases were traced back to the imbalanced dataset used for training, which primarily featured specic race or gender categories. As a result, the AI model struggled to accurately classify in the case of underrepresented groups. In addition to data selection bias, a models goals and validation metrics might impose biases on the model. Each of these biases can negatively impact marginalized and underrepre­sented groups, resulting in an unfair AI model.
To mitigate data selection bias and ensure that AI models in healthcare are fair and effective, it is crucial to train them on large and diverse datasets that accurately represent the patient population. This diversity needs to encompass various social statuses, minority groups, age groups, genders, races, and other relevant factors. By training on such datasets, AI models can learn to make decisions that are more inclusive and representative of the population they are intended to serve. This approach not only allows to reduce bias but also ensures that the resulting models are more robust and applicable across different patient demographics.
1.6.3 Data privacy violation
Modern DL methods heavily depend on large crowd-sourced datasets, which may contain sensitive or private information. Despite efforts to eliminate sensitive details, the presence of auxiliary knowledge and redundant encodings poses a risk of de­anonymizing datasets. Therefore, prioritizing a privacy-centric design is essential to safeguard individualsinformation, especially in applying DL techniques to critical
3
https://www.ml.cmu.edu/news/news-archive/2016–2020/2018/october/amazon-scraps-secret-articial-intel-
ligence-recruiting-engine-that-showed-biases-against-women.html
1-28
Artificial Intelligence in Adaptive Radiation Therapy
domains such as healthcare. Employing methods such as differential privacy and semantic security ensures data security throughout the model training process [69].
1.6.4 Risk and misuse
The aforementioned issues are primarily linked to poorly dened objectives and informational imbalances. However, even in cases where a system operates correctly, it can cause unethical conduct or deliberate misuse [70]. AI models trained on medical records might memorize patientspersonal information. This means that the AI could potentially recreate a patients data, raising privacy issues and making it difcult to ensure anonymity when using this data for research or development. In the worst-case scenario, AI memorizing patient data could be exploited by malicious actors, leading to situations where sensitive details are used for blackmail or other criminal purposes. Such models may also be misused by insurance companies to deny coverage or raise premiums based on pre-existing conditions gleaned from the memorized data. This could disproportionately harm patients with sensitive health histories.
In healthcare, the misuse of AI models can also have fatal consequences, such as intentionally manipulated AI algorithms used for medical diagnosis or treatment planning, leading to incorrect or harmful recommendations and patient outcomes [71]. Given the potential risks and misuse of AI applications, regulating these models seems necessary [72]. AI regulation refers to the establishment and enforcement of rules, standards, and guidelines governing the development, deployment, and use of AI systems. These regulations help to (i) ensure that AI systems are developed and used ethically and responsibly, (ii) safeguard individualsprivacy such as patients data, and (iii) establish accountability and liability in cases of AI-related incidents or harm.

1.7 Summary

This chapter provided a comprehensive overview of the foundational concepts of AI, focusing particularly on machine learning and neural networks. It covered the basics of machine learning, discussing various learning paradigms, such as supervised and unsupervised learning. The importance of feature engineering, linear separability, and classical models was emphasized, providing readers with a good understanding of the essential components of machine learning.
The chapter also explored articial neural networks, detailing the structure and function of different types of neural nets, including feed-forward, recurrent, convolutional networks, and transformers. Attention mechanisms and the training process for these networks were also covered. Applications and use cases of deep learning were highlighted, illustrating the practical impact of these technologies. Additionally, the chapter addressed model training and evaluation, discussing hyperparameters, data splits, evaluation metrics, and the common challenges of over-tting and under-tting. Finally, we touched on advanced topics such as generative models and their applications, and concluded with a critical examination
1-29
Artificial Intelligence in Adaptive Radiation Therapy
of ethical considerations in AI, such as transparency, bias, data privacy, and the risks of misuse.

References

[1] Russell S and Norvig P 2011 Articial Intelligence: A Modern Approach (Englewood Cliffs,
NJ: Prentice-Hall)
[2] Shortliffe E H 1974 MYCIN: a rule-based computer program for advising physicians
regarding antimicrobial therapy selection PhD Thesis Stanford University, Stanford, CA [3] Zadeh L A 1965 Fuzzy sets Inf. Control [4] Holland J H 1975 Adaptation in Natural and Articial Systems (Ann Arbor, MI: University
of Michigan Press) [5] Wells D M, Walrath D and Craighead P S 20000 Improvement in tangential breast planning
efciency using a knowledge-based expert system Med. Dosim. [6] Shortliffe E H, Scott A C, Bischoff M B, Campbell A B, van Melle W and Jacobs C D 1981
ONCOCIN: an expert system for oncology protocol management Proc. IJCAI 81 876–81 [7] Roberfroid B, Lee J A, Geets X, Sterpin E and Barragán-Montero A M 2024 DIVE-ART: a
tool to guide clinicians towards dosimetrically informed volume editions of automatically
segmented volumes in adaptive radiation therapy Radiother. Oncol. [8] Goodfellow I, Bengio Y, Courville A and Bengio Y 2016 Deep Learning vol 1 (Cambridge,
MA: MIT Press) [9] Bishop C M 2006 Pattern Recognition and Machine Learning (Information Science and
Statistics) (Berlin: Springer)
[10] Wold S, Esbensen K and Geladi P 1987 Principal component analysis Chemometr. Intell.
Lab. Syst.
[11] Ballard D H 1987 Modular learning in neural networks AAAI 647 279–84 [12] Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A and
Bengio Y 2014 Generative adversarial nets Advances in Neural Information Processing
Systems vol 27 (Red Hook, NY: Curran Associates) pp 2672–80
[13] Noroozi M and Favaro P 2016 Unsupervised learning of visual representations by solving
jigsaw puzzles European Conf. on Computer Vision (Berlin: Springer) pp 69–84
[14] Gidaris S, Singh P and Komodakis N 2018 Unsupervised representation learning by
predicting image rotations arXiv: 1803.07728v1
[15] Gal T S, Tucker T C, Gangopadhyay A and Chen Z 2014 A data recipient centered de-
identication method to retain statistical attributes J. Biomed. Inform.
[16] Theodoridis S and Koutroumbas K 2009 Pattern Recognition (Amsterdam: Elsevier) [17] Pedregosa F et al 2011 Scikit-learn: machine learning in Python J. Mach. Learn. Res. 12
2825–30
[18] Duda R O et al 2006 Pattern Classication (New York: Wiley) [19] Zhang-Salomons J and Salomons G 2015 Determine the therapeutic role of radiotherapy in
administrative data: a data mining approach BMC Med. Res. Method.
[20] Tan P-N, Steinbach M, Karpatne A and Kumar V 2016 Introduction to Data Mining (New
York: Pearson Education)
[21] Bagherzadeh P and Sadoghi Yazdi H 2017 Label denoising based on Bayesian aggregation
Int. J. Mach. Learn. Cybern.
[22] Parsa B 2022 Studies on decoupled modules for integration of extant knowledge sources PhD
Thesis Concordia University, Montreal
2 37–52
8 903–14
8 338–53
25 133–8
192 110108
50 32–45
15 1–9
1-30