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
Figure 9.10. Visual representation of the DSC.
pI I
,
()
XY
I, log
() ()
=II PII
XY
∑∑
XY
XY
⎜⎟
2
pI pI
()()
XY
,
where p is the probability distribution function of the intensities of X and Y, respectively, and p(I
, Iy) is the joint probability distribution function.
x
9.4.1.2 Feature-based registration
Feature-based registrations focus not on intensities between the two images, but instead on a de ned feature present between the two [41, 42]. These features could be points, lines, or entire structures, such as a contoured brainstem.
Feature-based registrations can broadly be described as minimizing the distance between points or surfaces of interest. When evaluating points, the sum of squared distances between the points appears very similar to the SSD equation shown previously:
N
oint differe nce
1
=−
()
PP
XY
=
i
1
ii
N
2
.
With a surface, this equation changes to evaluate the sum of differences from each point to that of the surface of interest.
A common metric of evaluating the overlap between two surfaces (2D or 3D) is the Dice similarity coefcient (DSC):
DSC 2 .
+
YX
XY
=
When two surfaces/volumes completely overlap, the DSC will be equal to 1, and when there is no overlap present it will be equal to 0, gure 9.10.
9.4.1.3 AI-based feature registration
Both intensity and feature-based registrations as shown above can be roughly called pre-made/hand-crafted features, that have been selected over time because of their
9-14
Artificial Intelligence in Adaptive Radiation Therapy
success. There are a variety of other options available as well, such as Gabor lters [43] and hand-crafted features that a user identies to best match their respective images [4446].
One of the most common uses of convolutional neural networks is in asking the question Are these the best features to guide our registration?and Can the model gure out what features are most useful?.
9.4.2 Types of registrations
Image registration algorithms have three main components: similarity index, trans­formation algorithm, and optimization method. For comprehensive descriptions, the reader is referred to existing chapters and review articles; here we will summarize for the context of AI in adaptive radiotherapy. Transformation algorithms are broadly classied into rigid and deformable.
9.4.2.1 Rigid registration
Rigid registration is characterized by applying a series of translations and rotations to the image. There are three potential translations relating to the left–right, superior–inferior, and posterior–anterior directions of the image, and three potential rotations relating to the pitch, roll, and yaw. These values can be used to directly relate any point in one image to another via the following transformation matrix, where M
, M2, and M3are the three translations, and Tx, Ty, and Tzare the three
1
rotations, gure 9.11.
A particular thing to note from this transformation is that there is a single, direct relationship between every point in reference B to reference A.
9.4.2.2 Deformable registration
Deformable, or non-rigid registrations, can be broadly categorized as anything that falls outside of the framework of rigid registration. Rather than maintaining a linear and direct relationship from each point in one image space to another, deformable registration enables each voxel to have its own individual transformation. The transformation of a single voxel from one space to the next can be described as the deformation vector eld: a matrix of vectors in the x-, y-, and z-plane which describe how an individual voxel from image A transforms into image B. Common deformable registration algorithms used in medical applications include Demons [47] based algorithms, B-spline [48], and the nite element model (FEM) [49, 50].
One of the concerns in deformable registration is this lack of consistent trans­formation. There is no requirement that every voxel in image A be present in image
Figure 9.11. Transformation matrix describing how a point (x, y, z) transforms from image B to image A.
9-15
Artificial Intelligence in Adaptive Radiation Therapy
B, and vice versa. Deformable registrations often have very complex optimization issues, with difculty interpreting results on a voxel-by-voxel basis. For this reason, there are many considerations regarding the optimization and evaluation of deformable registration models that need to be considered before clinical use.
When interacting with non-rigid registrations it is equally important to under­stand the implications of the registration results. For example, bone is not an object we would expect to readily deform. Just because the evaluation metric is optimal does not mean the deformation is realistic or accurate. The user should not evaluate a registration based solely on the same method used to guide the registration, as there is an implicit bias towards a positive result. For example, if the registration is driven by contour matching, the DSC between the contours is expected to be very high, however, this does not promise accurate or physiologically reasonable registration within the contoured structure.

9.5 AI-based image registration

Within image registration in AI there are multiple groups: supervised learning, unsupervised learning, reinforcement learning, and generative adversarial networks (GANs). For all techniques, the metric and methodologies used to evaluate the nal registration is vital as an indication of model efcacy. For all these groups, feature extraction via convolutional neural networks (new or pre-trained) is the common bedrock that connects them all. The features extracted can then be fed into existing models (Demons [47, 51], histogram matching [46], etc) as methods of registration, or used as inputs for new models.
9.5.1 Supervised learning
Within supervised learning, a previously determined registration between two images exists and is dened as the correctanswer. This is often the case in image registration, although this now limits the models ability to learn based on the accuracy of the previous registration and will perpetuate any systematic errors present in the ground truth training data.
9.5.1.1 Convolutional neural network registrations
Convolutional neural networks (CNNs) operate by convolving a number of kernels (typically 3 × 3 or 5 × 5 matrix in a 2D image) to identify features present in the images being registered. One of the advantages of CNNs is that these convolutions are not computationally expensive, the same kernel is applied across the space. However, a disadvantage is that the kernel itself has a relatively small receptive eld. If we imagine a single convolutional kernel used to identify diagonal lines (an ear), shown in blue in gure 9.12, we can easily see that while the kernel might be able to identify the ear, the receptive eld size might not be large enough to see enough of the image to overcome a poor initialization of the two images.
This represents a very important concept in many deep learning image registra­tion programs: both image initialization and the overall receptive eld size of the model can have signicant impacts on the nal accuracy of the prediction. For this
9-16
Artificial Intelligence in Adaptive Radiation Therapy
Figure 9.12. (Left) Representation of two images with an applied shift. The blue box represents a kernel for identifying diagonal lines (ear kernel). (Middle) Because of a poor initialization point and a lack of a larger receptive eld, the model could incorrectly register the nal two images based on this feature as a local minimum, rather than the (right) ideal registration.
reason, networks often include several layers of convolution and pooling, extracting both ne and coarse image information to guide the nal registration.
9.5.1.2 Generative adversarial networks
Generative adversarial networks (GANs) exist as a combination of two networks: a decoder which attempts to produce a realistic registration/deformation based on the input and target, and an encoder which attempts to differentiate between real registrations and generated registrations. GANs have shown success in the creation of synthetic modality images (synthetic CT from MRI [6264], T1-weighted to T2­weighted and vice versa [65]) and applied towards the medical registration task [61].
9.5.2 Unsupervised learning
Rather than be forced to train on explicitly dened answers, unsupervised learning for image registration requires that the image, either in whole or in part, be reducible to some form of feature vector that represents the patch/image. This can take the form of independent component analysis (ICA) [66] or PCA [28]. The models driving goal then is to reduce a complicated space/patch (2D or 3D MRI of the brain for example) into the principal components/features that represent that space. The model can then try to reduce the separation between these found features.
9.5.2.1 Reinforcement learning
These architectures are often built with some aspect of long-short-term memory (LSTM), the goal of which is to make an informed decision based on previous historical information. In the context of image evaluations, the convolutional LSTM [67] is often seen [55]. This imagines a 3D scan as a series of 2D images, each highly correlated to the image immediately above and below. The largest benet seen by reinforcement learning, is the rewardfeedback. Other regression-based methods rely heavily on the image initialization point. Reinforcement learning provides the
9-17
Artificial Intelligence in Adaptive Radiation Therapy
model an action, with a predened rewardacting as a feedback mechanism. Hu et al have shown that a simple transformation matrix: rotation, scale, and trans­lation can be used as the driving action, with a landmark based error acting as the reward function [55].
9.5.3 Registration in ART
The architectures presented in table 9.1 represent only a small portion of the work performed in AI-based registrations. These rapid registrations can alleviate temporal burden of recontouring targets and organs at risk, which would preclude online or real-time ART. They can further enable plan adaption by calculating a pre­determined plan on daily anatomy, representing dose of the day[6870]. With propagated contours and dose, the users can create a new plan going forward, or summate the dose on the original plan [71].
9.5.4 Commonalities in architectures
There are several deep learning components that are often seen in both supervised and unsupervised training. Auto-encoders and stacked auto-encoders are often set as the bedrock of deep learning models in image registration. The goal of such components is to remove unnecessary information and identify the principal components of the image.
We have previously identied both feature- and intensity-based metrics that can drive the registration process. Several AI models have instead ipped the question to ask: Why not allow the model to gure out what features are important?
9.5.4.1 Pre-trained convolutional neural networks
A common requirement from these architectures is the ability to identify useful features to guide the registration process. These features/intensities can be local or global and can either be user dened or determined by the model itself.
When using pre-trained models, the user must take extra care to match the new image intensities to that of the images used for original training. Further explanation of utilizing pre-trained networks is elaborated on later in this chapter within section
9.6.

9.6 Image segmentation

9.6.1 Introduction: coloring by the numbers
In the creation of a 3D treatment plan, the voxel-wise identication, or segmenta­tion, of targets and normal tissues/organs-at-risk (OARs) is a vital part of the planning process, with signicant downstream repercussions [7274]. Across a number of treatment sites, there can be a large number of OARs and targets required for the treatment planning process, creating barriers to re-planning, adaptation, and dose accumulation efforts due to the signicant time required for manual segmentation. Providing rapid and accurate segmentations is a vital part of the overall workow in ART, particularly in the online and real-time setting.
9-18
extraction, learned features
fed into histogram matching
Artificial Intelligence in Adaptive Radiation Therapy
algorithm
deformation prediction
hierarchical learning
(PEHL) > facilitates 2D/
3D registration via local
feature extraction
layers
LSTM post convolutional
3D registration vector field
Demons iterative approach
network used for feature
extraction, warping
transform from feature
distances
(Continued)
Brain MRI CNN CNN used for feature
deformable registration of MR brain
images
CNN Pose estimation via
T1/T2 Brain MRI CNN Encoder–decoder for
Fluoroscopic video,
registration
x-ray and
transesophageal
echocardiography
probe
2D/3D registration
learning
Reinforcement
nasopharyngeal
carcinoma
patients
Paired CT/MR
registration via reinforcement learning
SPREAD CT [57] CNN Pair of 3D patches generates
Brain MRI CNN 2D affine > 3D affine with
scale 3D convolutional neural networks
DR and DRR CNN Xception [59] pre-trained
diffeomorphic image registration with
very large deformations
via common representations learning
and differentiable geometric constraints
Table 9.1. List of various authors and techniques for rigid and deformable image registration. This list is by no means exhaustive but represents a small sampling of the
work done by various groups.
Author (year) Title Dataset Technique Explanation
Wu et al [52] Unsupervised deep feature learning for
Yang et al [53] Fast predictive multimodal image
Miao et al [54] A CNN regression approach for real-time
9-19
Hu et al [55] End-to-end multimodal image
Sokooti et al [56] Non-rigid image registration using multi-
Zhao et al [51] Deep adaptive log-Demons:
Liu et al [58] Multimodal medical image registration
from 3D patches, decoder
Auto-encoders identify features
network provides larger
FOV context
Artificial Intelligence in Adaptive Radiation Therapy
from GAN network,
evaluated based on Dice
of two cubic patches,
evaluated based on Dice
a dose of the day
registration between daily
CBCT and planning CT
structure propagation in
male pelvis from CT to daily
Algorithm comparison of
CBCT
auto-encoders
Brain MRI CNN with stacked
registration framework by
unsupervised deep feature
GAN Deformation fields generated
Retinal, cardiac
representation learning
CNN CNN estimating dissimilarity
images
using generative adversarial networks
A deep metric for multimodal registration Neonatal brain
MRI
Adult brain MRI
CT-CBCT Not AI, Nifty-Reg Evaluating dose summation as
and neck patients: Feasibility study on
using CT-to-CBCT deformable
CT-CBCT Not AI, normalized
registration for dose of the day
calculations
Deformable image registration for
gradient field
measure distance
and smoothing
adaptive radiotherapy with guaranteed
local rigidity constraints
regularizer
Table 9.1. (Continued )
Author (year) Title Dataset Technique Explanation
Wu et al [60] Scalable high-performance image
Mahapatra et al [61] Deformable medical image registration
(2016) [112]
Simonovsky et al
Veiga et al [70] Toward adaptive radiotherapy for head
(2016 [113]
Konig et al
9-20
Artificial Intelligence in Adaptive Radiation Therapy
As discussed previously, rigid, and deformable registration solutions offer opportunities for contour propagation from previous imaging. However, these methodologies all rely on the accuracy of the guiding registration and the accuracy of the previously dened structures. Strategies of identifying a subset of OARs based on disease site proximal to the target has also shown success in online ART [9].
Deep learning, particularly convolutional neural networks, have been shown to be highly successful in the task of OAR segmentation. For OARs, there are multiple vendor solutions available (MIM Contour Protégé [75], Raystation Deep Learning Segmentation [76], Radformation AutoContour [77], and Varian AI-Rad Companion [78]). Likewise, several groups have successfully created models for the segmentation of OARs present in the brain [79], head and neck [8083], lung [84
86], abdomen [87–90], and pelvis [91–94].
Target delineations offer several new difculties when creating predictive models: inter- and interobserver variabilities tend to be larger for targets compared to OARs [7274, 9598]. Particularly for clinical target volumes (CTVs) ‘a volume encom­passing visible gross tumor volume and subclinical malignant disease’ per ICRU 50 [99], there is often discussion about how generous to make certain contours, which leads to signicant challenges in automating this process.
Convolutional networks of CTVs and gross tumor volumes (GTVs) have been successfully implemented in a number of sites for brain tumors [100], rectal cancer [101], nasopharyngeal cancer [102, 103], breast cancer [104], oropharyngeal cancer [105], and arteriovenous malformations [106].
It is important to note that that with supervised machine learning, there is an implicit bias of the created model towards the training data. Models trained on manual contours from institution X could create suboptimal contours for institution Y if there is a systematic difference in practice between institutions. Furthermore, quantitative metrics do not always provide adequate clinical relevance of generated segmentations [94, 107109]. Including a qualitative assessment of the models predictions is vital for clinical feasibility.
9.6.2 Segmentation networks
A small sampling of convolutional neural networks tasked with the segmentation of OARs and targets is presented in table 9.2. While these studies all have unique qualities speci
cally designed to address the task at hand, there are several elements present across a majority of the networks. First is the difculty of both small eld high resolution and large eld context. As discussed previously in section 9.3, convolutions suffer from local dependence. Pooling layers can alleviate this depend­ence by reducing the overall search space for successive convolutions. However, as demonstrated in gure 9.4, recovering to the original image resolution after consecutive pooling layers results in a severe loss of ne resolution information. For this reason, many segmentation architectures incorporate skip connections. The most famous example of this is the fully convolutional neural network called U-Net [110]. Here high-resolution information from each part of the encoding path is directly transferable to the decoding path (gure 9.13).
9-21
Artificial Intelligence in Adaptive Radiation Therapy
segmentation
OAR
skip
connections
multi-head
attention
3D U-Net with
Mindboggle-101
Brain MRI
segmentation
OAR
ARTIX HN CT 3D organ-specific
residual U-Net
> nn-U-Net
OAR
HN CT Mirada medical
segmentation
segmentation
CNN
U-Net OAR
breast
Left-sided whole
MR/CT
segmentation
U-Net-GAN OAR
thoracic
challenge CT
2017 AAPM
enhancing
Tumor core,
> dense
Cascaded CNN
BRATS 2018
Brain MR
tumor, edema
segmentation
GTV, high-risk
connection
CT HN Stacked auto-
CTV
segmentation
encoder
Table 9.2. List of various authors and techniques for the segmentation of OARs and targets. This list is by no means exhaustive, but represents a small sampling of the
work done by various groups.
Author (year) Title Dataset Technique Goal
Segmentation using self-attention modules in MRI
images.
Deep 3D neural network for brain structures.
[79]
Laiton-Bonadiez et al
Cubero et al [80] Deep learning-based segmentation of head and neck
OARs with clinical partially labeled data.
organs at risk by deep learning contouring.
Van Dijk et al [83] Improving automatic delineation for head and neck
for improved cardiac sparing.
Morris et al [84] Cardiac substructure segmentation with deep learning
9-22
images using U-Net-GAN.
Dong et al [88] Automatic multiorgan segmentation in thorax CT
Ranjbarzadeh et al [100] Brain tumor segmentation based on deep learning and an
attention mechanism using MRI multi-modality brain
images.
risk oropharyngeal clinical target volumes with built-
Cardenas et al [105] Deep learning algorithm for auto-delineation of high-
in Dice similarity coefficient parameter optimization
function.
Artificial Intelligence in Adaptive Radiation Therapy
Figure 9.13. Basic representation of the U-Net style architecture, called such because of its Ushape. Note that information from the encoding (left) side of the architecture is maintained to the decoding (right) side, facilitating ner resolution segmentation.
Several alternative strategies exist, although at the fundamental level the goal stays the same: enabling ne resolution evaluation for voxel-wise segmentation, while gaining coarse resolution information for guided context.
9.6.2.1 Pre-trained convolutional neural networks
Pre-trained classication networks (VGG-16 [30], Xception [111], InceptionV3) can equally be applied for the task of semantic segmentation. Just as the features which are being extracted can be applied to other tasks in classication, the early convolutional layers can be useful in training a new segmentation model. Users often apply skip connections to the architecture encoder, freezing the previously trained layers and specically training an entirely new decoder. After an initial learning process, the original encoder layers can be unfrozenfor ne tuning, gure 9.14.
9.6.3 Best practices
9.6.3.1 Preprocessing
9.6.3.1.1 Intensity values
A major consideration for any convolutional neural network is data pre-processing. It is always benecial to normalize input images about the intensity values of interest. For example, if trying to segment the lungs, thresholding the HUs about values present in the lungs can help the model focus on important regions.
This window/leveling and thresholding is especially important when utilizing pre-trained networks. Recall that many pre-trained networks are trained on
9-23