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_980_Библиотеки_им_академика_М_И_Перельмана

.pdf
Скачиваний:
0
Добавлен:
31.08.2026
Размер:
28 Мб
Скачать
Brief Introduction toArticial Intelligence andMachine Learning
https://t.me/medicina_free
269
prostate classication at MRI [11], breast density at mammography [12], COVID-19 detection [13], and ventricle segmentation at cardiac MRI [14].
In this chapter, following a brief history of AI, and machine learning, we intro­duced how machine learning models work. Furthermore, we show how they can be applied in clinical studies.
2 A Brief History ofArtificial Intelligence
In 1955 John McCarthy, an assistant professor at Dartmouth College at the time, and his colleagues coined the term “articial intelligence” in their paper proposal submitted to a summer conference that was soon to be held at Dartmouth [15]. Although this event is believed to be the birth of AI as a distinguished subject, ini­tially it was Alan Turing, an English mathematician and computer scientist, who in his paper “On Computable Numbers” in 1936 proposed the concept of a computer [16] and then 14 years later in a 1950 paper titled “Computing Machinery and Intelligence” set the foundation of one of the rst intelligent machines [17]. For his pioneering efforts and contributions to this matter, he is now known as the “Father of AI” [18].
However, if we consider AI as a mighty tree grown from a seed that Turing planted and nurtured to a seedling by McCarthy and his collaborators, perhaps it is safe to say that Aristotle plowed the eld thousands of years before. The analy­sis of humans’ train of thoughts and how they perceive their surrounding world, discussed in Aristotle’s logic, was the philosophical foundation of what we now know as “computing” [19]; if we can identify a human brain’s method of think­ing, then perhaps we can create an “articial” brain that can mimic human intelligence.
Therefore, all different denitions that have been presented to explain AI derive from one similar concept: systems that can solve a problem by simulating the pro­cess of thinking and deduction [20]. McCarthy etal. wrote in their proposal:
The study is to proceed based on the conjecture that every aspect of learning or any other
feature of intelligence can in principle be so precisely described that a machine can be made
to simulate it. An attempt will be made to nd how to make machines use language, form
abstractions and concepts, solve kinds of problems now reserved for humans, and improve
themselves. [21]
After the Dartmouth summer conference, AI became one of the most popular elds in technology. Many books and papers were published concerning various types of intelligent systems. Two main theories were discussed in the AI eld: sym- bolic AI and connectionist AI [22].
The concept of symbolic AI is centered around the idea of encoding human knowledge into a system so that it can solve a problem based on the information given and conducting if-then-else rules. This school of thought came from McCarthy’s hypothesis, which was previously mentioned. On the other hand, Frank Rosenblatt introduced the concept of what we now know as connectionism. He
270
https://t.me/medicina_free
proposed that an intelligent system should be able to function like a brain: via neural networks (NNs). Over time, this system can recognize specic traits or patterns from analyzing the input data and, unlike symbolic systems, does not solve the problems solely based on preprogramming or traditional computer logic [23]. Rosenblatt named the articial neuron “perceptron.”
S. R. Motamedian et al.
2.1 Artificial Intelligence andMedicine
Soon, AI found its way into medicine. Even though the rst research in this eld began in the late 1950s, the main breakthroughs occurred two to three decades later with the development of unique expert systems such as MYCIN [24]—an intelligent system developed to help clinicians with diagnosis and antibiotics selection in patients with bacterial infections. An expert system is constructed through the implementation of symbolic AI algorithms, where the “inference engine” analyzes the input data according to the information available in the “knowledge base,” makes a decision, and presents it as the output [18]. Aside from assisting in diagnosis and therapy, automatic identication of potential adverse drug events and notication of patients and their healthcare providers of upcoming monitoring or preventive procedures are other tasks in which expert systems have aided clinicians [25]. Nowadays, most expert systems and generally symbolic AI have become obsolete. Instead, the main focus is on systems which can learn from provided data. In this case, there is no necessity for manually encoding the base knowledge into the systems. Machine learning is a branch of AI, dened as the study of developing intelligent systems that can learn by “gath­ering experience” [26]. In other words, the system receives the input data, ana­lyzes the patterns and structures concerning the desired task, and improves its performance by learning along the way. NNs, which started with Rosenblatt’s single neuron (perceptron), are considered a subgroup of machine learning. Progressions in the eld of deep learning started in the 1980s and more complex systems were developed by connecting several perceptrons to build a network [27]. One of the rst applications of NNs in medicine was William G. Baxt’s research in 1991, which used an articial neural network (ANN) for diagnosing acute myocardial infarction [28]. Currently, deep learning algorithms are widely used in the medicinal elds due to their remarkable ability in visual tasks such as object detection and recognition [29].
2.2 Artificial Intelligence andDentistry
From managing appointments and checkup reminders to diagnostic and treatment­planning assistants, AI has helped dental clinicians perform simple to complex tasks with higher efciency; it is faster and more persistent and needs less human resources [30]. Similar to the rst attempts on automation in medicine, AI made its way into
Brief Introduction toArticial Intelligence andMachine Learning
https://t.me/medicina_free
dentistry through expert systems. In a review published in 1992, Stheeman etal. reported ve expert systems in the dental eld [31]. The latest dentistry- related expert system at the time was “Oral Radiograph Diagnosis” presented by White in 1989 [32]. This program contained information about 97 conditions and was designed to assist in oral radiograph differential diagnosis using analytical decision-making. The other four expert systems reviewed in the article were used for treatment plan­ning [33], diagnosis of pulpal disease [34], orthodontic treatment advice [35], and oral diagnosis [36]. Nowadays, the growing fascination with automation has led to much research around developing various AI systems—mostly based on machine learning and, in particular, NNs—to assist oral healthcare providers in different den­tistry subelds. Some examples are artifact reduction in low-dose dental cone beam computed tomography (CBCT) in radiology [37], prognosis prediction of periodon­tally compromised teeth [38], skeletal age determination [39], caries detection [40], working length determination of the root [41], and prediction of bisphosphonate­related osteonecrosis occurring after tooth extraction [42]. Diagnosing osteoporosis [43], detecting TMJ osteoarthritis [44], evaluating periapical lesions [45], predicting oral cancer and locoregional recurrences [46, 47], and determining whether extrac­tions are needed prior to orthodontic treatment [48] are other examples of AI applica­tion in the oral and dental health eld.
271
3 How Does Machine Learning Work
3.1 The Learning Process
Generally, a machine learning model tries to learn the hidden pattern data itself or relevance to a label (e.g., reference standard). Regarding the type of learning that the machine algorithm is working on (and how it is trained), four main subtypes exist: (1) supervised, (2) unsupervised, (3) semi-supervised, and (4) reinforcement learning (Fig.2). A supervised learning algorithm relies on labeled data during the training phase. The algorithm is usually validated using assumptions that are based on human expertise. Typically, the algorithm tries to learn an input-output mapping so that inputs can be handled and outputs can be generated. Machine learning refers to “ground truth” as considered accurate and authentic data. Learning algorithms in unsupervised learning do not receive specied labels regarding the data. Data are clustered or grouped when machine learning is used to discover hidden patterns in the data. Semi-supervised learning combines both unsupervised and supervised techniques by combining labeled and unlabeled data.
Considering reinforcement learning, machines and computers carry out tasks and receive negative or positive feedback regarding the reinforcement process. Explicit instruction is not necessary to learn from reinforcement. Learning occurs as a result of interactions with the environment. Machine learning algorithms might be a mixed approach consisting of supervised, unsupervised, and reinforcement learn­ing [4956].
272
https://t.me/medicina_free
S. R. Motamedian et al.
Supervised Learning
Training
Set
Dog
Labels
Unsupervised Learning
Training
Set
Semi-supervised Learning
Training
Set
Dog
Unlabeled
Reinforcement Learning
Cat
Unupervised
Cat
Reinforcement
Deep Model
Model
Test Set
Semi-Supervised
Model
Actions
Rewards
Supervised
Cluster 1
Test Set
Environment
Model
Test Set
Dog
Cluster 2
Cat
Run?
Jump?
Fig. 2 Various types of supervisions in machine learning and their learning process
This is a Dog
This belongs
to cluster 1
This is a Dog
3.2 What Can WeDo Using Machine Learning?
Six main applications of deep learning in the context of healthcare are listed below (Fig.3):
1. Image processing: techniques applied either on raw signals or reconstructed
images (e.g., noise and artifact reduction) [57].
2. Object detection: identifying a particular subregion in an image, using bounding
boxes, expected to contain a specic object, for example, detection of brain tumors in magnetic resonance imaging [58].
3. Segmentation: extraction (or delineation) of a lesion or organ of interest based
on image analysis (e.g., pixel intensity, texture, edges), for example, vessel seg­mentation in coronary arteries to detect probable stenosis [59].
4. Classication: binary or multi-class categorization of a group or type from one
class to another. For example, wound classication in patients with chronic wounds is crucial for diagnosis that can be automated using deep learning [60].
5. Prediction: anticipating the progression of pathology. For example, in patients
who have cancer, this task may include predicting response to chemotherapy and the recurrence-free disease prognosis in treated patients [61].
6. Landmark detection: identifying desired landmarks on medical images. For
example, cephalometry landmark detection using deep learning algorithms showed a favorable performance in the literature [29].
Brief Introduction toArticial Intelligence andMachine Learning
https://t.me/medicina_free
Image Processing Object Detection
Segmentation Classficiation
273
Dog?
Cat?
Prediction Landmark Detection
Fig. 3 Various machine learning tasks that can be formulated as a medical problem for an articial intelligence-based approach
Dog
Cat
3.3 Data Collection andPreparation
Data preparation, including curation, de-identication, and labeling images for AI algorithm development, is necessary for having reliable outcomes. There are main steps that need to be taken to use medical image data in an AI algorithm. The rst is
274
https://t.me/medicina_free
local ethical committee approval by the institutional review board. Retrospective studies do not require the patients’ written consent, and additional procedures are not necessary. However, in prospective studies, informed consent is required. After ethical approval, data needs to be accurately de-identied and securely stored. Structuring the data in homogenized and machine-readable formats is the next vital step to be done [62]. The nal step is to associate the images with provided ground truth information, which can be one or more labels.
S. R. Motamedian et al.
3.3.1 Data Collection andCuration
Data collection and curation are steps performed to standardize and improve dataset quality for training deep NNs. Exploring and cleaning data is the most time-con­suming step in an AI project but is critical to any model training. By following general guidelines and human supervision, one can achieve high-quality standards in datasets building [63]. Data exploration evaluates the initial raw dataset’s general qualitative or quantitative properties to determine special features, global trends, or outliers.
3.3.2 De-identification
Due to the importance of patient-related privacy information, three concepts of de­identication, anonymization, and pseudonymization should be paid attention to. By de-identication, patient-related information will be masked from individual records [64]. However, by anonymization, the preferred approach for sharing medi­cal data, patient-related information from individual records will be removed irre­versibly. In pseudonymization, patient- related information will be substituted with synthetic values so that the original data can only be accessed with a predened secure key [65]. In addition to pixel information, digital imaging and communica­tions in medicine les need to be anonymized following protected health informa­tion regulations [66].
3.3.3 Resampling Medical Images
Medical images’ spatial resolution is generally higher than other image types. Besides, many medical image studies are three-dimensional instead of two-dimen­sional. So, computers with strong computational power are inevitable. Solutions to these problems are downsampling of the image resolution and patch-based analysis of image parts with relevant information.
3.3.4 Data Labeling andReference Standard
After selecting appropriate images, clinicians will classify samples, put landmarks, or delineate the region of interest through bounding boxes or segmentation masks. The reference standard, also identied as “ground truth,” must be dened before the training process. The term ground truth indicates information acquired from direct observation. Setting the reference standard may vary depending on the project. There are four approaches suggested:
1. Using gold standard: Consider we wanted to classify oral lesions based on intra-
oral photography to malignant and benign lesions. In this case, we can use histo-
Brief Introduction toArticial Intelligence andMachine Learning
https://t.me/medicina_free
pathology results for the reference standard. However, this approach is not applicable in most studies.
2. Consensus: A group of clinicians reaches a consensus on a diagnosis.
3. Major voting: A group of clinicians vote for the diagnosis. The label with the
most vote will be set as the reference standard.
4. The intersection of areas (only for segmentation studies): Two or more clinicians
draw segmentation masks, independently. Pixel-wise intersection of the segmen­tation mask area will be considered as the reference standard.
Various tools have been introduced to both the research and industry community. LabelMe and ITK-SNAP are suggested for 2D and 3D images, respectively. Both of these tools are open-source and free.
275
3.3.5 Dataset Sampling Strategies
The ability of an algorithm to achieve satisfactory outcomes on unseen data is called generalization. The entire dataset needs to be divided into different sets to increase generalization. The samples in all groups should be independent of each other and identically distributed. The most adapted sampling strategy in deep learning is dividing the dataset into training, validation, and test sets. The optimal ratio varies in different situations, but as a rule of thumb, the proportion of 80% train, 10% validation, and 10% test division is suitable. Using the validation set, optimal hyperparameters are selected. Ultimately, the algorithm is used once on the test set to measure nal performance. k-fold cross-validation is another com­monly used sampling strategy used in small datasets. Data is split to k equally folds. The algorithm is trained on n-1 folds in each training step and tested on the left fold. The mean of this repeated k-times process gives the nal performance. However, this sampling strategy cannot be deployed in a large dataset because of computational cost.
3.3.6 Challenges ofSmall Datasets
The main challenge in applying deep learning in the clinical eld is inadequate data samples since small sample sizes hinder generalizing the model to unknown data [67]. Since building large medical datasets of excellent quality is challeng­ing and costly, some innovative training strategies and model architectures have been proposed, such as weak labeling [68] or few shots learning [69] to address these limitations, especially in the medical eld. However, the primary approach to confronting this challenge is using transfer learning. Transfer learning is related to applying an earlier-learned approach from a task to another but a simi­lar project. Using transfer learning, we can use labeled data from one set of tasks to do another. It provides a promising opportunity for machine learning to advance rapidly in different domains. For instance, a convolutional neural net­work (CNN) for classication and pattern recognition of nonmedical images on ImageNet (http://www.image- net.org) has been applied to chest CT scans to extract their features and predict the survival rate of lung tumors [70, 71]. Accordingly, medical image analysis can benet from knowledge gained from nonmedical image analysis.
276
ab
cd
https://t.me/medicina_free
S. R. Motamedian et al.
3.4 Model Implementation
3.4.1 Machine Learning Models
There are various numbers of machine learning models used to solve different prob­lems. Here, we briey introduce some of the most important ones (Fig.4).
Linear Regression
The linear regression technique is probably one of the most widely used and com­prehensive statistical and machine learning methods. When looking for a linear rela­tionship between two variables, linear regression is often utilized. Using linear
X
+
Y
Fig. 4 Selected machine learning techniques. (a) Logistic regression. The likelihood of a binary result is modeled using logistic regression. The Y axis in the image indicates probability, whereas the X axis represents the continuous input variable. Take note that minor changes in X result in huge differences in the nal probability Y, particularly in the middle region of the gure, where the model’s uncertainty is greater. This approach may be extended to issues with many classes. (b) Support vector machine. This model may convert a nonlinear border to a linear one. The training procedure maximizes the distance between classes and the nal chosen border. (c) Random forest. Random forest is a strategy for minimizing the uncertainty in the nal forecast by combining deci­sion trees. It is based on a recursive binary splitting technique in which the higher nodes are meant to be the most discriminative and subsequent branching is applied to variables that are less impor­tant. (d) Clustering. Clustering is a method that enables the identication of subgroups (clusters) within a set of data. There are several clustering algorithms; some need a predened number of clusters, others may be employed with output data, and still others are completely unsupervised
Brief Introduction toArticial Intelligence andMachine Learning
https://t.me/medicina_free
277
regression, there are two options: single and multiple. A model with a single inde­pendent variable is known as a simple linear regression [72]. Using a simple linear regression, we can see how a variable varies. Simplied regression separates the effects of independent variables from the interactions of dependent variables [73]. Multivariate linear regression (MLR) is a statistical approach that uses several explanatory factors to predict the outcome of an answer variable. The purpose of MLR is to simulate the linear connection between the independent variables x and y [74].
Logistic Regression
Contrary to a binary classier, logistic regression (LR) models rely on cumulative input variables to estimate their degree of predictability. This model can have a disproportionate impact on nal probability predictions in a slight change in the input value. Furthermore, the number of predictor variables should be limited since it might cause escalating training process, overtting, and consequent weak model robustness. The LR model is a suitable alternative for binary classication of less complex data.
Support Vector Machine
Support vector machines (SVMs) employ supervised machine learning techniques to divide data into distinct categories by determining the optimal boundary. Gaussian function (or linear function) is the most commonly utilized kernel. Regarding the pros and cons, SVM models need vast storage when applied to a high amount of data; however, SNM would be a model of choice to identify nonlinearity and spar­sity in the data.
Random Forest
A random forest (RF) model is an ensemble of decision trees (DTs) trained on ran­dom samples taken from a training set. Decision trees are a set of explicit rules adjusted for classifying various features of the training set. Moreover, DTs are known as high variance nonlinear models. Accordingly, very deep and accurate DTs tend to reveal irregularities and overtting issues. Consequently, applying the RF model to several datasets may solve this issue. Therefore, the model is getting more generalized and more applicable. The mode for classication goals and the mean for regression analysis goals are the most optimal choice of obtained predictions. Notably, two main parameters should be considered: (1) the number of applied deci­sion trees and (2) the depth level of each decision tree. It is worth mentioning that the more depth levels, the more inuential the model would be, and in contrast, the less generalized the model would be.
Cluster Analysis
Cluster analysis is dened as grouping data points based on similarities and prox­imities in the quantitative analysis of variables extracted from input data to obtain the space of the parameters. Cluster analysis may help in patient stratication (e.g., based on pathology images). Moreover, k-nearest neighbors (kNN) is the most
278
https://t.me/medicina_free
S. R. Motamedian et al.
common clustering model used for classication in a supervised way. In nding subgroups, k is the number of neighbors to be considered. Accordingly, predictions for a new patient are based on its surroundings.
3.4.2 Deep Learning Models
Various structures have been suggested for deep learning models, mostly based on ANN.Here, we briey introduce some of the most well-known structures and mod­els (Fig.5).
Artificial Neural Networks
Biological NNs inspire ANN patterns. ANN models generate predictions from input data in a hierarchical way, using nodes and layers. An ANN can adapt until it is found that the weights (i.e., input lines) give the best performance. In the case of large amounts of data, ANNs are the best choice. In this section, we present basic neural network architectures.
Convolutional Neural Network (CNN)
CNNs are the most common deep learning algorithms used for image analysis. CNNs have shown successful performance measures regarding various image analysis tasks (e.g., detection, classication, segmentation). As mentioned above, hidden layers are a part of CNNs, which are functional layers (i.e., convolutional layers, pooling layers, and/or fully connected layers) converting the input data to output. In simple NNs, the neurons of one specic layer are connected to every neuron of the previous layer; thus, the layer is considered fully connected. However, a deep NN architecture consisting of fully connected layers is very computation­ally expensive, and tting a large number of parameters is challenging [2]. In CNNs, the neurons are only locally connected. The pattern used to calculate the connections for one neuron is reused over all neurons of the same layer [75]. Therefore, CNN has drastically fewer parameters than a fully connected deep NN of the same size.
Recurrent Neural Networks (RNNs)
Recurrent neural networks (RNNs) are another subset of NNs mainly used for sequential data (e.g., MRI and ultrasound image sequences). RNNs can “remem­ber” the past and use the knowledge derived from the past in making current deci­sions. Considering the function of RNNs, the rst image is used as an input in this method, and predictions are made based on that. Further, RNNs memorize the pre­vious predictions and predict the following image sequence. RNNs widely contrib­ute to model long-term memory. In this regard, LTSM [76] and gated recurrent unit (GRU) [77] are the most popular.
Generative Adversarial Nets (GANs)
Goodfellow etal. [78] advocated using a generative adversarial network (GAN) for picture synthesis from noise. Generative articial neural networks (GANs) learn to