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_117_библиотеки_им_акад_М_И_Перельмана

.pdf
Скачиваний:
2
Добавлен:
15.09.2026
Размер:
15 Мб
Скачать
☆
230
compared to other occupations, it was found to cause higher pain in the neck and back. Specically, it was linked to higher levels of prolonged low-intensity muscle activity, less rest time for the trapezius muscle, and more time spent doing nonsurgi­cal activities and the two functions in RAS.Patient activation was above average during open surgery.
For robot-assisted surgery (RAS), Peng etal. [28] suggested an endoscope (FOV) autonomous tracking method that takes into account pose control, and image deni­tion, allowing the AEHR system to reliably and swiftly follow, which incorporates robot-assisted endoscope holding, was developed in light of clinical needs and the limits of conventional research. Second, limitations on the RCM of AEHRs and the corresponding kinematic model equations are constructed. After using deep learn­ing to segment and nd tool tips, an intelligent adjustment control model for the surgical FOV is constructed, complete with a wide variety of indices. The project concludes with the development of a prototype system for experimenting with AEHRs in conjunction with human discourse.
R. S. R. Somula et al.
3 Methods
In order to model a robotic surgical job for estimate purposes, we rst establish a surgical estimation nite state model (FSM), and then we create the surgical HFSM, which models a full RAS operation, an inoperable estimate. Due to the complexity and great dynamic range of the real-world RAS situation, the FSM model does not impose any necessary constraints on the probability of any state transitions. Figure1 is a block schematic of the suggested model.
Fuzzy Sets and Membership Function
Fuzzy sets and membership functions are concepts from fuzzy logic, a mathemati­cal framework used to handle uncertainty and vagueness in data and decision­making. Fuzzy logic is particularly useful in cases where traditional binary logic (true/false) is inadequate due to the presence of intermediate or partially true values. In traditional fuzzy set theory, a fuzzy set is dened by a membership function that assigns a membership degree (a value between 0 and 1) to each element of the uni­verse of discourse. The membership function determines the degree of belonging­ness of each element to the fuzzy set.
Hierarchy Construction
The HFSM extends the concept of fuzzy sets to a hierarchical structure. The rst step is to construct a hierarchy of fuzzy sets, organized in a tree-like structure. Each node in the hierarchy represents a fuzzy set at a specic level of abstraction or granularity. The root node represents the most general fuzzy set, and subsequent levels represent more specic and detailed fuzzy sets.
Explainable AI (XAI)-Based Robot-Assisted Surgical Classication Procedure
Fig. 1 Decay, transfer, and compose (DTC) perfect for the nding of surgical state estimation
231
Multilevel Membership
In HFSM, an element can belong to multiple hierarchical fuzzy sets simultaneously. This is in contrast to traditional fuzzy sets, where an element can belong to only one fuzzy set. Each hierarchical fuzzy set at different levels has its membership function that determines the degree of membership for an element at that level.
Membership Degree Propagation
One of the key features of HFSM is the ability to propagate membership degrees from parent nodes to child nodes in the hierarchy. This means that the membership degree of an element in a child fuzzy set is inuenced by its membership in the par­ent fuzzy set. This allows for a more exible and accurate representation of complex relationships between elements.
Aggregation and Composition
HFSM allows for the aggregation and composition of membership values from dif­ferent levels of the hierarchy. This process combines membership degrees from par­ent and child nodes to obtain a more comprehensive representation of an element’s membership in the overall system. Various aggregation and composition methods can be used depending on the specic application.
Granularity Control
HFSM provides the ability to control the level of granularity at which the data is analyzed. By selecting different levels in the hierarchy, users can focus on different levels of detail and abstraction, making the model adaptable to various analysis requirements.
Applications
The hierarchical fuzzy sets model nds applications in various domains such as decision recognition, information retrieval, and expert systems. It is particularly useful in handling complex and uncertain data where multiple levels of abstraction need to be considered.
232
Handling uncertainty: Fuzzy logic allows for the manipulation of uncertain or
imprecise information, which is common in real-world scenarios. Human-like reasoning: Fuzzy logic can mimic human reasoning, which often
involves degrees of truth and vagueness. Flexibility: Fuzzy logic can be applied in various domains, including control sys-
tems, decision-making, pattern recognition, and articial intelligence,
among others. Interpretability: The use of linguistic variables and membership functions makes
fuzzy logic models more interpretable and understandable for humans. Smooth transitions: Fuzzy logic enables smooth transitions between different states,
which can be useful in control systems.
Fuzzy logic and its concepts have found applications in areas such as industrial control systems, expert systems, robotics, and more, where uncertainty and impreci­sion are inherent to the problem at hand.
Denition 1 What is the nite state model (FSM) for surgical estimation?
M(S,Σ,F) is consisting of S, an innitely many non-empty state space, the sym­bols sent into the system, and P, a set of allowed changes among conditions.
Denition 2 Requiring in-patient surgery, a surgical FSM with states that are potentially other surgical FSMs is called a hierarchical FSM (HFSM). An HFSM used in surgery has superstates for its states.
R. S. R. Somula et al.
Denition 3 A superstate is the largest possible breakdown of an operation into discrete stages.
These denitions just scratch the surface of a full (hierarchical) nite state machine model; the dynamics of transitioning between states is outside the scope of this chapter. Therefore, an HFSM is analogous to a graph modeling the connections between levels of organization. In this chapter, we ignore the possible inuence of banned transitions on estimator performance because our study does not contain any surgical conditions with such restrictions.
Using an input data stream, the hierarchical surgical state estimator (i.e., DTC) estimates operations that may include a more complex hierarchy, and we will keep things basic for now. The next section discusses the structure and instruction of DTC.
3.1 Datasets
Data from 20 actual transabdominal preperitoneal repairs of inguinal hernias done using da Vinci®Xi surgical systems are included in the HERNIA-20 dataset [11]. There are 12 single-sided and eight double-sided inguinal hernia repairs in HERNIA-20. The eight steps in the process represent the eight superstates in an HFSM.We proceeded to subdivide the ve superstates (SS0–SS4) into a total of
Explainable AI (XAI)-Based Robot-Assisted Surgical Classication Procedure
four intermediate states and one nal ne-grained surgical state. Superstates like this showed up often in HERNIA-20 and are of great therapeutic signicance. Because repeated states were classied as the similar class in training and assess­ment of the at approximation techniques that we test against, overlaps of ne­grained states exist among distinct superstates. Surgical texts and interviews with practicing surgeons were used to determine (super)state designations. When esti­mating a state, DTC places no restrictions on the probability of transitions between states.
A HERNIA-20 operation might take anywhere from 15 to 70min. There are two types of laparoscopic instruments. A da Vinci endoscope is used to acquire visual data during an endoscopic procedure. Four universal patient-side manipulators (USMs) are used to gather the robot kinematics data. Ten system events are recorded: six are binary of console (endoscope follow, instrument follow, surgeon-side con­sole master clutches, energy pedal) and four are categorical (da Vinci® surgical tool put on USMs). Because of the thorough anonymization of HERNIA-20 data, the name of the surgeon(s) performing each surgery is concealed.
233
3.2 DTC Method
The approach recommended for identifying HFSM is described in depth here. The workow and the formalization of the technique are discussed, beginning with an impression of the architecture and moving nished the various mechanisms of the approach.
3.2.1 DTC Architecture Impression
DTC model contains three steps. In the rst step, we will use the DTC’s pretrained CNN perfect as our backbone to train a model to excerpt deep local characteristics from respectively image. The next step is to use DTC’s class-decomposition layer to streamline the local structure of the data. In the second step, the training is con­ducted using a complex gradient descent optimization algorithm. Finally, a class conguration layer is used to further enhance the nal picture categorization of the study. Knowledge is transformed from an ImageNet pretrained CNN model and added. Each class in the picture dataset is broken down into k subclasses in the class decomposition layer so that they may be processed separately. Next, the class com­position module is used to reassemble the subclasses into a concluding classica­tion of the initial picture collection.
234
A
nn nm
12
⋯
ll
{}
=−
i
ac
R. S. R. Somula et al.
3.2.2 Deep Feature Extraction
(Using the HERNIA-20 dataset, a pretrained CNN perfect was adapted and trained with a shallow-tuning mode. The image feature space was built using the commer­cially available CNN features of pretrained, which is completed just on the nal classication layer.) Images have a high dimensionality; therefore, we projected the high-dimensional feature space into a lower-dimensional space using PCA [29]. This allowed us to remove characteristics that were highly linked. This is a crucial part of the class decomposition process that leads to more uniform classes, lessens the framework’s memory footprint, and boosts its overall performance. In transfer learning, a pretrained CNN model is used as a starting point, and additional layers are added to the model to ne-tune it for a new task. The large dataset for a general vision-related task is like image classication on ImageNet. However, the learned features in the early layers of the pretrained model can be valuable for various vision tasks, even if the target domain is different.
To leverage transfer learning, you can add new layers, often referred to as “top layers,” to the pretrained model. These new layers are typically task-specic layers responsible for adapting the features learned by the pretrained model to the new task. For example, if you want to use the pretrained model for a different image classication task, you may add fully connected layers and a new output layer for the specic number of classes in your target dataset.
3.2.3 Class Decomposition Layer
Now imagine that our eye space is signied by a 2D matrix and L is a class group­ing. The letters A and L may be rearranged to form
aa a
11 11 1
 
aa a
=
21 22 2
⋮⋮
 
aa a
⋯
⋯
m
 
Ll
,,l
m
⋮
⋮
 
,,…
=
1
k2
(1)
The total number of features is m, and there are k classes and n images. Using k-means clustering [30–33] for class decomposition, we then divided each class into similar subclasses (or clusters). Here, the squared Euclidean distance (SED) is used to assign each pattern in the initial class L to the class label of the centroid with the smallest SED.
n
k
SED
∑
=
where the centers of each group being represented by the letters cj.
j
j
()
∑
i
1
=
1
j
(2)
AALBBC=
()
→=
()
||
al
al
11
21
bl
bl
mc
11
21
Explainable AI (XAI)-Based Robot-Assisted Surgical Classication Procedure
235
After the clustering is complete, a new dataset will be formed by subdividing each class in L into k equal subclasses (denoted as dataset). Accordingly, a mathe­matical description of the connection among datasets A and B may be expressed as:
(3)
If A and B each have the same sum of instances and C stores the updated names of the classes (e.g., C={l_11,l_12,...,l_1k,l_21,l_22,...,l_2k,...l_ck}), then A and B can be rewritten as:
 
aa
11 11
aa
21 22
 
⋮⋮⋮⋮⋮
A
=
  
aa all
12
nn nm
 
⋯
⋯
⋮
⋯
 
bb
11 11
 
bb
21 22

B
=
⋮⋮⋮⋮⋮
  
bb bll
nn nm c
12
 
⋯
⋯
⋮
⋯
m
m
⋮⋮⋮
m
⋮⋮⋮
      
2
2
 
 
1
    
21
 
2
 
(4)
3.2.4 Transfer Learning
It is likely that the various classes will be adequately represented due to the abun­dance of large-scale annotated picture datasets. This means that the boundaries between classes that are learnt are more likely to apply to novel data. On the other hand, the generalization error may grow due to the scarcity of annotated HERNIA-20 data, especially when severely affected than others in terms of size and illustration. This is due to the possibility of a misalignment between the demographics of the minority and the majority groups. When tens of millions of parameters need to be trained, large-scale annotated datasets (such as ImageNet) give viable solutions through transfer learning.
In the experimental research part, we will look at how we employed and con­trasted multiple ImageNet modes for transfer learning. Overtting can arise when using stochastic gradient descent (SGD) with insufcient training data because of
236
E
n
j
n
()
()
()
=
0
R. S. R. Somula et al.
the large uctuations in the objective/loss function. Minimizing the objective func­tion E(·) using cross-entropy loss helped in convergence and prevented overtting:
jj
yzx
,
()
=−
1
jj
1111
ynzx yn zx
∑
+−
()
()
j
−
j
 
(5)
where xj are the training set of input photos, yj are the true labels, and z(·) is the projected output of a SoftMax function.
3.2.5 Procedural Steps ofDTC Model
After going through the theoretical underpinnings of the DTC model, we will move on to a demonstration and summary of the DT model’s operational phases in Algorithm 1.
Algorithm 1 Procedural Steps of DTC
1:Procedure
–Input –Imagesetdivided into training and testingset –Ground−truth labels. –Output –Predicted labels.
Stage1Class Decomposition:
2. Usean ImageNet pretrainedCNNmodel(e. g., AlexNet)asafeature extractor to constructadeep feature extractor to constructadeep feature space from input images.
3. ApplyPCAonthe deep features space for dimension reduction.
4. Usereduced feature space of the input images to decompose original classes intoanumber of decomposed classes.
Stage2Transfer Learning
5. Adapt the nal classication layer of an ImageNet pretrainedCNNmodel to the decomposed classes.
6. Fine−tune parameter of the adopted pretrainedCNNmodel.
Stage3Class Composition
7. Calculate the predicted labels associated to decomposed classes.
8. Rene the nal classication using error−correction criteria.
4 Results andDiscussions
The recommended model was developed using Python 3.7, NumPy 1.16.2, NetworkX 2.4, and TensorFlow 2.1.0. The tests were run on a computer with 32GB of RAM, a seventh-generation Intel Core i5 processor, and an 8GB GeForce RTX 2070 graphics card.
A
TP FP TN FN
+
++ +
Pr
TP FP=+
Recal
+
F
+
..
Explainable AI (XAI)-Based Robot-Assisted Surgical Classication Procedure
237
4.1 Evaluating Models
Accuracy used to assess the projected models (F1). Eqs. (6)–(9) are the formula used to regulate these values.
ccuracy
ecision
12=
Precicion Recall
TP TN
=
TP
TP
l
=
TP FN
Precicion Recall
(6)
(7)
(8)
(9)
Table 1 presents the validation analysis of proposed prototypical with existing techniques.
Table 1 represents the performance of cross-validation performance and test per­formance. In this analysis, we computed the different model and proposed method; the experimental analysis of proposed model delivers the better accuracy and other performances. For instance, the accuracy of proposed model for CV is 98% and the same model achieved 92.22% of accuracy for TP. In addition, the nearly 98% of PRE, REC, and F1 for CV, where the DTC accomplished nearly 92% of PRE, REC, and F1 for TP. However, the existing techniques achieved nearly 88–96% of accu­racy, PRE, REC, and F1 for CV and these techniques achieved nearly 79–89% of accuracy, PRE, REC, and F1. When compared with all techniques, DT and CNN models achieved low performance, i.e., nearly 74–76% of accuracy (PRE, REC, and F1) for CV performance. Figures2, 3, 4 and 5 present the graphical comparison.
Table 2 represents the HERNIA-20’s surgical superstate approximation perfor­mance of DTC.In this analysis, we computed the different models such as Trivial,
Table 1 Comparative analysis in terms of cross-validation with test performance
Approaches Cross-validation (CV) presentation Test presentation (TP)
BiLSTM 96.96 97.0 97.0 97.0 90.89 90.9 91.89 91.89 RNN 94.89 94.87 94.87 94.87 88.6 88.61 88.6 88.6 DTC 98.08 98.09 98.08 98.08 92.22 92.23 92.22 92.22 DT 78.95 82.7 78.95 78.34 67.91 68.95 67.91 67.43 CNN 74.76 78.02 74.76 74.03 64.05 66.79 64.05 62.45 LSTM 94.92 94.92 94.92 94.92 90.18 90.18 90.18 90.17 RF 87.51 87.63 87.51 87.5 86.09 86.13 86.09 86.09 SVM 86.15 86.67 86.15 86.1 78.97 79.98 78.97 78.78
ACC PRE REC F1 ACC PRE REC F1
238
Fig. 2 Accuracy comparison
R. S. R. Somula et al.
Fig. 3 Precision assessment
ImageNet, Robot kinematics, and raw endoscopic vision with proposed method; the experimental analysis of proposed model delivers the better result of ned tuned.
Explainable AI (XAI)-Based Robot-Assisted Surgical Classication Procedure
Fig. 4 Recall assessment
239
Fig. 5 F1-score assessment
5 Conclusions
In this study, we offer a straightforward description of a RAS procedure as a multi­level time-varying state-space model (HFSM) used in surgery. A surgical HFSM has several superstates, each of which is an FSM made up of discrete surgical