Добавил:
Sekretar
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_537_Библиотеки_им_академика_М_И_Перельмана
.pdf
factor–related challenges can be faced upon manual measurements that
include errors and associated manual calculations and results being observer
dependent. An android-based smartphone application called Facekit is
developed and tested for clinical settings in [1] using a head tracking device
and an ML Kit. The cellular application is tested on 15 patients for digital
and direct measurement methods’ comparison by collecting orbital,
orbitonasal, naso-oral, and naso-facial ratio data. The authors concluded
that Facekit’s highest intraclass correlation accuracy of the two methods
was 0.32, which is very low compared to the benchmark value of 0.75,
which is considered to be an excellent agreement between the digital
method and manual method. Even though the results attained are weak, this
first version of Facekit can be the first step toward further advanced
versions in the future.
Convolutional neural networks (CNN) as deep learning is applied for
wrinkle detection in the literature for plastic surgical purposes. Facial
feature recognition prediction depends on wrinkles on the forehead, corner
and under-the-eye regions, and cheeks, while forehead wrinkles have been
used frequently for age estimation [29]. This is due to possible age
progression indicators such as skin texture, face structure, and skin color,
noting that face features change with age progression of a human. Wrinkles
that are present in the cheek region that relate to the curvilinear nature of
the nasolabial line, which extends from the nose to the mouth region on
both left and right sides, are used in age analysis [30, 31]. Figure 1 displays
the age progression of a child using a computer algorithm.
Fig. 1 Ageing progression of a child using an algorithm up to age 80 with only frontal view [30]
Complete face detection method is utilized in [2] using CNN to
determine wrinkles that can be easily identified even if some parts of the
face are not recognized by the algorithm. A classifier is developed and used
by the authors to reduce weaknesses for outcome advancement. Efficient
identification of skin flaws was the key for analysis of wrinkle prediction;
insufficient skin flaws are eliminated in the facial region of interest by
https://t.me/medicina_free

analyzing at what stage the wrinkles appear. The elements that are
determined to be weak are removed from the associated region. A dataset of
Kaggle data source [32] containing a total of 754 facial images with and
without wrinkles is used by the authors with the associated CNN training
and testing sets developed. The Hough transformation method is used to
process images for detecting shapes with their mathematical
representations. The designed CNN contained two hidden layers. Upon
testing using Python IDE, a wrinkle detection accuracy of 95.85% is
attained [2].
Facial anthropometry data and analysis of such data have been used for
plastic surgery applications and the design of facial protective equipment.
Facial morphology requires a large amount of associated morphology data
based on the 3D facial specs, and the database classification is
accomplished by using the measured facial parameters. The classic method
of 3D points’ selection on facial landmarks is through manual observations
and the associated manual calculations upon patient interaction.
Advancement of this classic method is accomplished by using deep learning
algorithms with automatic point selection/marking and measurement of the
facial features. Fuzzy clustering method is a way to classify data points in
which each data is allowed to be a member of different clusters; Fuzzy
clustering algorithmic analysis can be used for classification and counting,
and calculations related to the facial data according to the morphological
key facial points can be accomplished along with the length calculations
including nose width, face width, and eye width. Effectiveness of the
algorithmic approach in comparison to classic method is conducted by
comparing the two methods’ outcomes through different statistical analysis,
such as regression or correlation analysis with error terms’ determination. A
CNN model that has such a classification is designed to obtain facial feature
points containing a convolution layer, a pooling layer, and a fully connected
layer in [3]. Rectified linear unit activation function is used by the authors
for fully connected layers and L2 loss function is used for optimizing the
entire model for attaining the coordinate positions of the facial points. For
the application of the CNN on a dataset, image enhancement using rotation
and translation is applied on 5000 facial images containing manually
marked 10 key points: four of the points associated with the tip, middle, and
corners of the nose; four points attained from left and right corners of the
eyes; and two points attained from the corners of the mouth. Using DL for
https://t.me/medicina_free

automatic facial measurements and fuzzy clustering analysis of facial
images into 65 groups, the authors of [3] concluded that the developed
CNN had an average error rate of less than 0.8% and an average failure rate
of less than 4.5% in the feature point positioning when compared to other
existing algorithms.
A plastic surgery dataset is used for training, testing, and validating a
CNN developed in [14] in order to perform facial recognitions after plastic
surgeries by extracting features with the associated feed forward (input to
output progressing) model’s classification of both normal and surgically
altered images. Noting the computational complexity of calculations related
to the associated network’s architecture, it is paramount to reduce the GPU
and computational requirements for the training process along with the
minimal time usage to train the network over the CPU. A loss function can
be used to describe how much closer the current model gets to the accurate
model and helps in adjusting the parameters of the model to further
decrease the loss in next iteration. A loss function used by the authors
helped to drop the number of incorrect positives for each iteration to
increase the accuracy of the model, and such an accurate model reduces the
computational requirements and complexities of the system without
compromising the precision value that is linked to accuracy. Figure 2
displays the general structure of a CNN with multiple layers and their
image correspondences.
https://t.me/medicina_free

Fig. 2 CNN with multiple layers used for image processing and recognition [14]
The authors reported a training accuracy of 100% and validation
accuracy of 82% over plastic surgery images. From a theoretical framework
standpoint, the authors concluded that a CNN-based facial recognition
model can work with minimal GPU requirement by using a single CPU,
making the method of calculations more efficient. Establishment of a realtime robust identification system is needed.
Machine learning has been applied in ophthalmology and evaluation of
eyelid contours [34–38]. A deep learning-based image analysis to
automatically compare eyelid morphology corrections upon blepharoptosis
surgery using pre- and postoperative images in [39]. Generative Adversarial
Network (GAN) is used as a part of plastic surgical procedures in [4].
Two neural networks are generated and compete with each other in a
Generative Adversarial Network (GAN) in the machine learning theory to
improve accuracy of their predictions. In the associated GAN generation,
training and testing of the developed deep learning is based on the pre- and
postoperative blepharoptosis surgery–related facial image data collected
between 2016 and 2021 at an oculoplastic clinic. There are four modules:
data processing module, ocular detection module, analyzing module, and
prediction module, which are used in the designed postoperative appearance
prediction system with the GAN model. The goal of the authors was to
https://t.me/medicina_free

design an automated deep learning model that can accurately predict
postoperative appearance of blepharoptosis surgery patients. This approach
is expected to result in displaying patients their appearance after the surgery
and reduce their concerns. The absolute errors using predicted and actual
marginal reflex distances are calculated for determining the accuracy of the
model. Additionally, feedback is attained from 750 pairs consisting of
experts and patients, and the analysis indicated a high satisfaction rate of
56% and satisfaction rate of 35.7%.
A deep neural network (DNN) with two stages of cascaded fully
convolutional network (FCN) is developed for facial landmark detection in
[6] facial palsy from coarse to fine. In an FCN, only locally connected
layers (such as convolution and pooling) are used; avoiding the use of dense
layers allows the network to be trained faster due to the reduced number of
parameters. The authors of [6] developed a database containing annotated
facial landmarks for facial palsy for testing of the developed DNN.
Recognition of faces that experienced plastic surgery is conducted in
[11] by using a Deep Feed Forward Neural Network (DFFNN). The back
propagation of this developed neural network to update the weights of the
model is the key feature of such an approach to optimize the training ability
of the model with a reduction on the computational complexity of the
solution.
Bootstrapping sampling with a confidence interval of 95% is used for
application on the plastic surgery facial database. The authors used metrics
such as Recognition Rate, mean square error, F-score, and regression
coefficient for comparison purposes with the existing models. The
developed DFFNN results are determined to be equivalent to the existing
CNN-based models, such as Deepface (Alexnet), FaceNet (GoogleNet),
VGGface (VGGNet-16), Light CNN (Light CNN), SphereFace (ResNet-
64), Cosface (ResNet-64), Ring loss (ResNet-64), and Arcface (ResNet-
100). Recognition rate values of 98.2% obtained for local surgeries and
97.9% obtained for global surgeries are the best results attained when
compared to the commercial CNN methods mentioned above. Each facial
image of the 88 subjects in the database is independently and manually
annotated with 68 facial landmarks and 16 classes of asymmetric facial
expressions, including right eyebrow, left eyebrow, right smile, left smile,
right snarl, left snarl, right wink, left wink, gentle eye closure, tight eye
closure, brow raise, close smile, frown, funny, open smile, and snarl. The
https://t.me/medicina_free

authors of [11] used conventional methods such as Emotrics [39], an opensource software utilizing machine learning for automatic facial landmark
detection for patients with facial palsy, and Supervised Descent Method
(SDM) [40], a widely used algorithm for facial landmark detection.
Proposed DNN by the authors of [11] performed better than the Emotrics
and SDM based on the dataset developed by the authors; hence, landmark
comparison of normal and palsy faces indicated the importance of the
developed facial landmark database specifically for facial palsy for further
performance improvement.
From a technological standpoint, there are commercially available facial
recognition and analysis software AI-driven tools, including the following:
Amazon Rekognition by Amazon [41]
FaceMe AI Facial Recognition Engine by Cyberlink [42]
Google, Vision AI [44]
Luxand Face Recognition API [46]
Kairos face recognition [43]
Microsoft Azure Face Recognition API [45]
A commercial AI software offered by Haystack.ai is used in [8] for
assessment of 65 orthognathic surgery patients’ age and attractiveness
prediction. Authors of [8] tested Haystack’s AI solution for its ability to
match “attractiveness” of the postoperative orthognathic surgery patients
with the definition of attractiveness used for the developed software. The
AI application was significantly better than both pre- and postsurgical
ratings of the humans based on statistical calculations. Similarly,
attractiveness ratings of AI solution differed from the human raters but still
reflected the same general trends.
Deep learning is used in [17] to simulate postoperative facial
appearances of orthognathic surgery patients with jaw deformity
corrections. A facial shape change prediction network is introduced by the
authors for mapping bony facial shape changes to facial shape changes. Preand postoperative data are used in the network that utilized weak
supervision. Conventional methods in this area of interest mainly utilized
biomechanical modeling with Finite Element Method (FEM) in the
associated simulation, which can be labor intensive and computationally
challenging. The results attained by the authors in [17] are compared with
one of the conventional methods utilizing FEM, introduced in [47], called
https://t.me/medicina_free

realistic lip sliding effect (FEM-RLSE). Qualitative and quantitative
analysis of pre- and postoperative computed tomography (CT) images of 24
female and 16 male patients with jaw deformities are used for analysis and
comparison purposes to evaluate the proposed method with the
conventional methods.
The proposed network’s qualitative evaluation indicated its achievement
of 9 better predictions, 18 comparable predictions, and 13 worse predictions
when compared to FEM-RLSE. Wilcoxon signed-rank test used for the
statistical significance of the two methods with the proposed network to
designate three points for a better prediction, two points for a comparable
prediction, and one point for a worse prediction, indicating the proposed
and the conventional methods to be comparable up to an accuracy of
p>0.3. The proposed network model is determined to be 15 times faster
than the conventional FEM-based method.
Two machine learning models for facial deformation detection and
measurement are developed in [18]. One of the models is trained on a
dataset of 200,000 normal faces utilizing a convolutional autoencoder that
serves as a variant of CNN utilizing unsupervised learning of convolution
filters, and such an autoencoder is used for reconstruction of the database
images by learning optimal filters to minimize reconstruction errors; this
reconstruction error attained from the autoencoder is used as an indicator of
facial abnormality in [18]. The other model is trained on the same 200,000
normal faces and additional images that entirely lacked normality with the
face detector confidence score serving for measurable outcomes on the
results. Both ML results are compared with the human subjects’ evaluations
on a 7-point scale. The evaluation results on 80 persons’ evaluations of 60
images indicated high correlations between their average score and the ML
results. The authors concluded the possibility of using the two ML models
as a possible handheld tool for facial deformation detection.
Aesthetics is one of the areas that machine learning is applied in [53].
An automated classifier is designed using supervised learning, and the
training of the model is completed by using 165 facial features extracted
from attractive female face images that were also independently graded by
human referees. A variety of descriptive attributes that relate to
postoperative target variables such as facial ratios are collected as the
attributes to be used in the decision trees developed. The results were
strongly dependent on the human classification of facial beauty. High
https://t.me/medicina_free

accuracy of automated classifier made it a possibility to be used in the
future as a predictive tool for estimating a patient’s perceived beauty
following aesthetic surgery, providing a quantitative measure to possibly
discourage patients from undergoing procedures that offer marginal
improvement by setting expectations.
3 Fracture Detection Through Machine Learning
Applications
There is a shortage of specialists and plastic surgeons that can diagnose
facial bone fractures that motivated researchers to develop ML applications
that can fulfill such tasks [9]. There are limited number of studies that relate
to the fracture detection on the facial bone structure due to several
limitations and challenges. The first difficulty that a machine learning
application faces is to read and classify facial bone fractures due to its
occurrence in many different places in different forms. The second obstacle
is the challenge faced by the architect of the ML application for reading the
facial structure and designing an ML solution that can fulfill training and
testing accuracy expectations. For instance, in the case of applying ML to
cervical bones, it is rather easy to successfully classify and predict
nonuniformity of the cervical bone by using ML due to its uniform
structure. There are limited number of ML solutions developed for facial
bone fractures and we will cover the associated literature in this section.
The diagnosis of nasal fractures incorporates human factors related
challenges and can result in a variability of readings and results as well as
the time it takes to determine the fractures; this can be particularly critical
in the diagnosis of traumatic nasal fractures from facial CT analysis
perspective for early reconstruction. Due to this reason, the ability to detect
nasal fractures automatically by using deep learning is investigated in [5]. A
3D-CNN structure is used by the authors for attaining the goal by
incorporating facial CT scans and a 3D computer-aided diagnostic system.
In this design, the computer-aided diagnostic system served as the
automatic diagnostic system due to its ability to detect various types of
fractures in the literature, such as hip and vertebral fractures [48–50]. The
experimental setup was based on the dataset consisting of 2535 patients’
nasal bone CT images, with 53.25% normal and 46.75% fractured bones,
https://t.me/medicina_free

and binary nasal bone classification is conducted to distinguish between
normal and fractured bones using the 3D-CNN. 3D-ResNet-34 and 3DResNet-50 are used as deep learning models. Training, testing, and
validation of both normal and fractured nose data are designed to scale the
learning ability and evaluation of the deep learning model. The learning of
the 3D-CNN plays a significant role in the outcomes, and a 3D-CNN model
has automatic feature extraction and selection with spatial dimensioning
during the training of the model to optimize performance. Using wellknown calculations for the evaluation of DL models, including the area
under receiving operator curve, sensitivity, specificity, and accuracy, the 3D
voxel data approach with cubic topology achieved the highest performance
during the binary 3D classification of nasal bone based on the facial CT
using a single network. In this outcome, 150 errors occurred on average and
both models reported errors due to minor fractures without depressed
fractures or deviated nose. The approach followed in [5] indicated the
possibility of effective use of AI for automatic diagnosis of nasal bone
fractures, with 3D-CNNs in a facial 3D CT image voxel dataset.
Another deep learning application for fracture detection on CT images
is developed in [9]. The authors used an object detection model called
YOLOX-S that is trained by using the Intersection over Union (IoU) loss
value; a value that varies between 0 and 1 specifying the overlap percentage
amount between the predicted bounding box of the image by the algorithm
and the ground truth bounding box. In the case when IoU=1, there is 100%
overlap between the bounding box used by the algorithm and the ground
truth box, and a value of zero means there is no overlap between the boxes.
In addition, the authors modified the data augmentation methods of
YOLOX and introduced a new data augmentation method [52] to maximize
the effect of data augmentation. The test dataset consisted of CT scans of 40
patients; 47.5% of these images were fracture free attained from normal
patient, 5% of the images had both nasal bone fractures and other facial
fractures, 2.5% containing only other facial fractures and no nasal bone
fractures, and the remaining scans containing only nasal bone fractures. The
training CT scan data are selected only from the nasal bone fractures. The
measured instances included average precision, sensitivity, specificity, and
F1 score. Among the 14 models generated by using 7 different loss
functions, one of the systems is observed to be good at classifying whether
a person has a fracture or not; however, average precision is determined to
https://t.me/medicina_free

be only 69.8%. We refer to [9] for further details on the developed models
and the best model generated.
4 Use of Deep Learning for Preoperative Facial
Simulation
Postoperative face simulation, diagnostics, and patients’ risk stratification
are designed by using deep learning in [10]. The authors initially used an
adaptive median filter that allows discrimination of pixels in the filtering
window as corrupted and uncorrupted depending on the nature of the pixel
determined through statistical calculation with a follow-up filtering
technique applied to corrupted pixels in the window. The median value of
the pixels in the filtering window is used for replacement of the noisy
pixels. In addition, a computer image processing technique called
Histogram Equalization is used by integrating the Laplacian partial
differential equation approach to improve the contrast in the images. A
model called Smart restorative frustum model is used for converting into
3D.
The deep spatial Multiband VGG NET CNN is used by the authors for
postoperative face prediction by using approximately 310,000 CT scans
with the associated clinical records attained from different clinical centers.
Two batches are formed by the authors to validate the algorithms clinically.
The authors observed Jaccard and dice scores to be successful metrics that
resulted in accurate outcomes compared to other traditional methods by
using MATLAB. Accuracy, specificity, and sensitivity values of 93.7%,
99.8%, and 99.9% are attained, respectively, by using the suggested
classifier and these values are determined to be higher than the same values
of the other known approaches. Therefore, the authors concluded that
postoperative face prediction is observed to be better than other existing
mechanisms.
Computer vision and deep learning have been integrated in several
studies with the goal of developing predictive models for blepharoplasty
morphology, craniosynostosis, orthognathic dysmorphisms, and facial burns
that may result in surgeries [13].
There are general surgical procedures developed that relate to burn
research that utilizes machine learning to classify burn depth and determine
https://t.me/medicina_free
Соседние файлы в папке Библиотека им академика М.И. Перельмана
