Читать книгу Digital Dentistry - Группа авторов - Страница 33

Artificial intelligence (AI) is a term commonly used to represent the concept of artificially created human‐like intelligence [11]. AI has revolutionized several fields of scientific research and development, with improvements ranging from facial recognition on our smartphones to genome analysis. AI has been developed in different sectors of society to solve various tasks by utilizing data more optimally, including health areas like medicine and dentistry. Data science is an umbrella term that refers to the interdisciplinary area dedicated to the analysis of structured and unstructured data, which aims at the application of statistical methods for data analytics. The work of data scientists includes using advanced techniques such as machine learning. Machine learning is a subfield of AI, in which algorithms are applied to learn patterns intrinsic to data structures, which allow data predictions (supervised machine learning) and data reduction and data handling (unsupervised machine learning) [12]. One of the factors associated with the adoption of this methodology would be that AI is especially suitable for overcoming the variability in subjective individual examination and for increasing the effectiveness of diagnosis [13].

Unsupervised machine learning is used for reducing the dimensionality of such datasets, increasing interpretability. Clustering uses metric variables as input to group data, enabling reduction of data into smaller groups, with categorical variables as outputs. Principal component analysis (PCA) also enables data reduction by grouping variables that correlate with each other. Multiple correspondence analysis (MCA) is used for data reduction and data handling of categorical variables that associate with each other. Unsupervised machine learning is very suitable for handling a large amount of data and for data mining, acquired through dental patients' records for example [14].

Supervised machine learning, defined by its use of labeled datasets, is used for prediction. Accuracy will therefore depend on the size and quality of the database, as well as the labeling and training method. Supervised machine learning uses linear regression, logistic regression, ensemble models, and neural networks (NNs). The main component of supervised machine learning is labeling the dataset to train the algorithms and to accurately classify data or predict results (Figure 1.4).

Neural networks are particularly useful for complex data structures, such as imaging data, as these models are capable of representing an image and its hierarchical resources, such as edges, corners, shapes, and macroscopic patterns [15, 16]. During the training process, the corresponding data and labels are transmitted repeatedly by the NNs. The computational power of these NNs depends on the quality and quantity of training data, which allow these networks to update the weights assigned to each variable of the model in question, and this training can be supervised or not. NNs can also be associated in several layers. The term deep learning is a reference to deep (multilayered) NN architectures [12].

Figure 1.4 LabelImg AI software tool being used to label the dataset.

Recently, convolutional neural networks (CNNs) have been developed and applied in many aspects of the health field, including dentistry, with the performance of various tasks like image classification and object detection [11, 12]. CNNs use a clever trick to reduce the amount of training data needed to detect objects under different conditions. The trick is basically to use the same input weights for multiple artificial neurons – so that all those neurons are activated by the same pattern – but with different input pixels. Every convolutional layer responds to stimuli only in a restricted region of the image's field of interest, known as the receptive field. This structure differs from conventional image classification algorithms and other deep learning algorithms, as the CNN can learn the type of filter created manually in conventional algorithms [15, 16].

When training image datasets are entered into a machine learning system, the learning procedures are automatically repeated, without the need for a manual definition of the image characteristics. In this way, machine learning methods (with or without deep learning) can adaptively learn image characteristics and simultaneously perform image classification [16]. Thus, the results from models that use machine learning differ greatly from systems that operate with conventional programming. From the moment that a NN is trained, weights are assigned to the characteristics of the sample data, and the results are intrinsically linked to the sample used for training this model. Therefore, more favorable results can only be found with AI training with a new, larger, and more accurate sample.

Methods based on machine learning depend on the quantity and quality of information that can be learned (based on a set of training data). A smaller sample size can reduce the potential for identification accuracy in test images, thus decreasing the sensitivity and specificity of the test. Furthermore, a large sample can be created by using data augmentation, which is a feature used to increase the sample of images by using image rotation and resizing. Thus, based on a limited number of data, it is possible to increase the sample size, which can help researchers to work with small samples and avoid overfitting. Overfitting occurs when the intrinsic data of a specific sample are prioritized; thus, the performance of the model for the images not belonging to the sample ends up being much lower than the performance related to the sample used for training.

Several authors have reported high precision using NNs for object detection in the classification of dental elements [17], periodontal disease [18], dental caries [19], apical lesions [20], cystic lesions and tumors [21, 22], dental fracture diagnoses [23], and sinusitis in the maxillary sinus [24], among other factors, with the use of deep CNNs applied to panoramic radiographs and digital periapical radiographs. However, although methods for object detection via deep CNNs are rapidly progressing, object detection can still be challenging, as it is dependent on a large database and computational processing [25, 26].

Object detection software using AI already exists. Second Opinion^®, for example (Hello Pearl, Los Angeles, USA), is an AI‐driven assistant for diagnostics and treatment planning. The software offers a computer vision platform that can instantly detect dozens of common pathologies (Figures 1.5–1.8).

In addition to object detection in radiographic analysis, NNs have been used for automatic identification. Examples include comparison between antemortem and postmortem panoramic radiographs (human postmortem identification) [25], diagnosis of osteoporosis on panoramic radiographs [26–28], and malocclusion diagnosis [29, 30].

Another important possibility with the use of CNNs is the automation of identification and selection of structures in CBCT DICOM images. With this method, automated mandibular canal detection [31] and automated mandible segmentation [32], an approach to dental implant planning [33], are possible.

Figure 1.5 Screen capture of the software Hello Pearl (Los Angeles, USA) showing automatic detection of a maxillary area with bone loss.

Figure 1.6 Screen capture of the software Hello Pearl (Los Angeles, USA) showing automatic detection of a mandibular area with bone loss.

Figure 1.7 Screen capture of the software Hello Pearl (Los Angeles, USA) showing automatic detection of marginal discrepancy of a metallic restoration.

Figure 1.8 Screen capture of the software Hello Pearl (Los Angeles, USA) showing automatic detection of calculus.