Journal Search Engine
Search Advanced Search Adode Reader(link)
Download PDF Export Citaion korean bibliography PMC previewer
ISSN : 1226-7155(Print)
ISSN : 2287-6618(Online)
International Journal of Oral Biology Vol.48 No.4 pp.45-49
DOI : https://doi.org/10.11620/IJOB.2023.48.4.45

The principles of artificial intelligence and its applications in dentistry

Yoohyun Lee, Seung-Ho Ohk*
Department of Oral Microbiology, School of Dentistry, Chonnam National University, Gwangju 61186, Republic of Korea
*Correspondence to: Seung-Ho Ohk, E-mail: shohk@chonnam.ac.kr https://orcid.org/0000-0001-9828-1143
November 30, 2023 December 12, 2023 December 12, 2023

Abstract


Digital dentistry has witnessed significant advancements in recent years, driven by extensive research following the introduction of cutting-edge technologies such as CAD/CAM and 3D oral scanners. Until now, 2D images obtained via x-ray or CT scans were critical to detect anomalies and for decision-making. This review describes the main principles and applications of supervised, unsupervised, and reinforcement learning in medical applications. In this context, we present a diverse range of artificial intelligence networks with potential applications in dentistry, accompanied by existing results in the field.



초록


    © The Korean Academy of Oral Biology. All rights reserved.

    This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/bync/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

    Introduction

    Long after the introduction of deep learning, which is the fundamental base of artificial intelligence (AI), it has become the leader in various industries, including the medical field. Recently, AI has aroused worries among medical specialists that AI might replace occupations such as diagnostic radiologists. However, after a few years, massive research has begun, leading to the understanding that it must not replace humans. The decision-making part is where AI could not take over; it is solely designated for humans. It is now convinced that AI, as a tool for supporting diagnosis, is effective. Several companies like VUNO Co. Ltd. or Lunit Inc. have appeared as leaders in this supportive AI. However, there is also a wide variety of research going on other than imaging. AI in the medical industry could be divided into several parts: diagnosis, treatment planning, and managing after treatment. Here, AI could provide over-human accuracy and efficiency. This could be the essential role of AI where the medical industry needs the help of hands, running out of time in most hospitals. Also, dentistry has its distinctiveness where the patient can also possess general illness, which needs advice from a physician. AI might work on those issues or distinctly on dentistry.

    Otherwise, deep learning could be categorized into the big 3 paradigms: Supervised, Unsupervised, and Reinforcement learning. Here, in this paper, we will talk about the 3 parts and which networks belong to each category. Within each domain, we will elucidate the distinct roles and notable accomplishments of individual networks, delineating the specific contributions and achievements of each. This will give a clear picture of how well each network performs and what contributions it brings to its field.

    Beyond diagnostics, deep learning plays a pivotal role in automating and enhancing treatment planning strategies. Tailored treatment plans, personalized to individual patient characteristics, are becoming increasingly achievable through the integration of ML algorithms. This review will explore how deep learning algorithms contribute to the optimization of treatment plans, particularly in areas such as implantology and surgical interventions. As dentistry evolves towards a more patient-centric approach, the role of deep learning in predicting treatment outcomes and optimizing therapeutic interventions becomes a focal point of discussion.

    Furthermore, this paper will address the implications of deep learning on the security and privacy of patient data in the dental domain. As the reliance on digital records and collaborative platforms increases, safeguarding sensitive information becomes paramount. An exploration of the current landscape of data security measures and potential avenues for improvement will be presented, ensuring a balanced understanding of the ethical considerations associated with the integration of deep learning in dental practice.

    In summary, this review aims to provide a broad perspective on how deep learning is making waves in dentistry. By examining current applications and looking ahead to future possibilities, we contribute to the ongoing conversation about how deep learning is transforming oral healthcare practices.

    Supervised Learning in Medical Applications

    Supervised machine learning (ML) is where ML has inputoutput pairs to train. There are known outcomes that should be achieved by the trained network. Therefore, the labeled datasets are trained as input and target. This leads networks to make predictions or decisions based on input data, guided by labeled examples where the correct outcomes are known.

    Each input is labeled with the corresponding ground truth or target, which is recently made by human experts or existing medical records. The next step is the training phase, where the labeled dataset is used to train the supervised learning model. During training, the model learns to map input features to the corresponding target labels by adjusting its internal parameters. Model prediction is where, after trials of training and adjusting over and over, the model is put to make predictions or classifications on new, unseen data. It uses the learned patterns from the training data to infer the likely outcome for a given input. Lastly, evaluation and validation are also critical points. The model’s performance is evaluated on a separate set of data not used during training. Via validation or a test set, assessment of model training can be done.

    The outcomes of supervised learning can be discrete, called a classification algorithm, where it can be processed as continuous values. The classification model is processed as discrete values or classes. In this algorithm, according to separate thresholds, they will be categorized. Prediction of continuous value is a regression algorithm. In medical AI, a regression algorithm is used to predict numeric outcomes, such as estimating patient outcomes, predicting disease progression, or determining the optimal dosage of a medication. Linear regression models, which are used to linearly relate input features and target variables as a linear equation. Support vector machines (SVM) also aim to find a hyperplane that best captures the trend in the data while allowing for a margin of error.

    After the setting and training of the algorithm, adjusting features of the network is another task for successful ML. The network is trained with modified variables to reduce the error of prediction using an optimization technique. What usually happens here is the possibility of overfitting to a specific dataset, which jeopardizes the network’s ability to generalize. A significant amount of time can be put into optimizing the network. There are ways to minimize the risk of overfitting. First is to separate datasets into folds. Therefore, there are trained sets and validation sets that affect the learning of the next cycle. Lastly, the model is assessed by test sets (Fig. 1).

    1. Image analysis and interpretation

    This task is excessively studied worldwide. Using computed tomography (CT), magnetic resonance imaging (MRI), x-ray images, the networks are trained to identify and classify features or abnormalities in medical images. They can classify or segment where the target features are. Also, it is applicable when deciding the position to place dental implants. Convo-lutional Neural Networks (CNNs) are often used as the main structure for this kind of tasks. Consisted of convolutional layers that can automatically and adaptively learn spatial hierarchies of features. They are widely used in image classification, object detection, and other computer vision tasks. Tooth-Net (Fig. 2) uses mask regions with convolutional neural network (Mask R-CNN) (Fig. 3) by Facebook inspired network to segment each tooth from cone-beam computed tomography (CBCT) image and produce 3D whole dental arch [1,2]. Here, U-Nets have been playing an extensive role in medical image segmentation. Sivagami et al. [3] has done dental panoramic image segmentation based on U-Net. Most recently there is CNN-transformer architecture U-Net for CBCT image segmentation [4].

    2. Disease prediction and risk stratification

    Using supervised learning could be applied when estimating how patients will respond to specific treatments based on various factors, including genetics and clinical parameters. Disease prediction is done after training with input data; unseen data is put into the network, and the output label indicates the presence or absence of a particular disease, which can be other than disease depending on the data used. Therefore, the goal is to build a predictive model.

    Unsupervised Learning Approaches in Medical Research

    Unsupervised learning is a type of ML where input data does not require explicit output data. This leads the algorithm to explore inherent features within data rather than features set by human interpretation. The main goal is often to discover the underlying relationships, structures, or distributions within the data. Among various types, medical data often possesses high-dimensional and complex features, making unsupervised learning techniques valuable for discovering hidden structures or patterns within such data. This learning could be particularly useful when revealing patterns of data that may not be apparent through manual inspection. However, inexplicit algorithmic results may not convince actual users, so unsupervised learning remains a rich area for ongoing exploration and research.

    1. Clustering is used when grouping data

    Some features or tendencies of the illness can lead to data clustering. Autoencoders are used to detect anomalies. The architecture consists of an encoder and a decoder, where the encoder compresses the input data into a lower-dimensional representation, and the decoder reconstructs the input from the encoder. The goal is to minimize the reconstruction error. Training the autoencoders is carried out with normal data, such as patient records, images, or time series data. The reconstruction error is detected by threshold, and detection is done. Anomaly detection can be integrated into clinical decisionsupporting systems, providing healthcare professionals with timely alerts and aiding in decision-making.

    2. Feature extraction and dimensionality reduction

    Data such as MRI and CT, genomics, and electronic health records are high-dimensional data. Using high-dimensional data means they can be computationally expensive, making it challenging to analyze and interpret effectively. To work with a reduced set of features to speed up training, dimensionality reduction is used. High-dimensional datasets are susceptible to overfitting, where a model performs well on training data but poorly on unseen data. Dimension reduction might help reduce the risk of capturing noise in the data. Unsupervised learning includes Principal Component Analysis (PCA), t-distributed stochastic neighbor embeddiing (t-SNE), Uniform Manifold Approximation and Projection (UMA), Sparse coding [5,6]. As seen above, these techniques might already be familiar as they belong to traditional statistics.

    3. Data augmentation

    Generative Adversarial Networks (GANs) are a typical example of unsupervised ML where they specialize in generating synthetic data aimed to produce the most realistic data. There are GANs that generate new data without any predetermined augmentation method [7]. Cycle-GAN was used to generate synthetic non-contrast CT images by learning the transformation of contrast to non-contrast CT images. Another type of GANs called deep convolution generative adversarial network or Conditional-GAN also exists; however, the challenge using GANs is that the newly generated images possess the same distribution characteristic as trained data, which means no new information is added. However, other GANs are under research for a more accurate, novel method.

    Reinforcement Learning in Personalized Treatment and Decision-Making

    Reinforcement learning (RL) is a branch of ML where learning to make decisions by interacting with an environment to achieve a goal. In the context of medical AI, RL has been applied to address complex tasks. In this paper, this is not our focus; however, we will simply discuss typical examples.

    1. Personalized treatment plans

    In this plan, optimizing treatment plans for individual patients can be planned by AI. This helps in the adaptation of treatment strategies based on the patient’s response to the learning process. As the process continues, the treatment outcomes can improve. Radiation therapy optimization is used to maximize treatment efficacy while minimizing damage to healthy tissues.

    2. Clinical decision support systems

    As RL algorithms learn from historical patient data and medical literature to provide personalized recommendations for diagnosis, treatment, and patient management. Clinical decision support systems (CDSS) aim to assist healthcare professionals in making informed decisions by providing relevant information and recommendations based on patient data and medical knowledge. While traditional CDSS often rely on rule-based systems or ML approaches, the integration of RL algorithms in CDSS is an emerging area with the potential to enhance decision-making in healthcare settings.

    Challenges and Future Directions

    1. Data privacy and ethical considerations

    As more AI is getting involved in the medical field, it is critical not to evade ethics, especially considering patient information. Unauthorized access to medical AI systems can lead to exposure of patient identities and their health data, which should be kept confidentially. Therefore, methods to control access, encryption, and authentication mechanisms are adapted.

    2. Interpretable models and explainability

    Medicine is a field of cause-effect where the reasons should be verified clearly to plan treatment. When applying the networks and interpreting the output from them, professionals should be interpretable for their decision-making process. More explainable networks will be required to convince practitioners.

    3. Future directions and integration

    This review is to bring AI closer to practitioners’ minds. Clinicians without previous ML involvement often leave out the choice of using AI due to its limitations. However, over a few years, it would be inevitable to follow the stream of AI being involved, which this paper prospects as an effective tool. One should not rely too much on the results produced, though, which makes the need for more papers interpreted by clinicians. It’s essential to note that ethical considerations, regulatory frameworks, and ongoing collaboration between AI developers and dental professionals are crucial for the successful integration of AI in dentistry. Additionally, developments in AI may lead to changes in the landscape, and staying updated with the latest research and industry trends is recommended.

    Conclusions

    In conclusion, the summary of key findings from this review highlights the significant contributions and challenges encountered in the application of ML to the medical field. This synthesis provides a comprehensive overview, emphasizing the importance of continued research, collaboration, and innovation as essential components for maximizing the benefits of ML in healthcare.

    Funding

    This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF- 2015R1D1A1A01057503).

    Conflicts of Interest

    No potential conflict of interest relevant to this article was reported.

    Figure

    IJOB-48-4-45_F1.gif

    Illustration of supervised learning.

    IJOB-48-4-45_F2.gif

    Tooth-Net.

    IJOB-48-4-45_F3.gif

    Mask regions with convolutional neural network (Mask R-CNN), a representation of most segmentation techniques available.

    Table

    Reference

    1. He K, Gkioxari G, Dollar P, Girshick R. Mask R-CNN. IEEE Trans Pattern Anal Mach Intell 2020;42:386-97.
    2. Cui ZM, Li CJ, Wang WP. ToothNet: automatic tooth instance segmentation and identification from cone beam CT images. Paper presented at: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2019 Jun 16-20; Long Beach, USA. Piscataway: IEEE, 2019. p. 6361-70.
    3. Sivagami S, Chitra P, Kailash GSR, Muralidharan SR. UNet architecture based dental panoramic image segmentation. Paper presented at: 2020 International Conference on Wireless Communications Signal Processing and Networking (WiSPNET); 2020 Aug 4-6; Chennai, India. Piscataway: IEEE, 2020. p. 187-91.
    4. Chen Z, Chen S, Hu F. CTA-UNet: CNN-transformer architecture UNet for dental CBCT images segmentation. Phys Med Biol 2023;68:175042.
    5. Blei DM, Ng AY, Jordan MI. Latent Dirichlet allocation. J Mach Learn Res 2003;3:993-1022.
    6. van der Maaten L, Hinton G. Visualizing data using t-SNE. J Mach Learn Res 2008;9:2579-605.
    7. Motamed S, Rogalla P, Khalvati F. Data augmentation using Generative Adversarial Networks (GANs) for GAN-based detection of Pneumonia and COVID-19 in chest X-ray images. Inform Med Unlocked 2021;27:100779.