Introduction
Oral cancer is one of the most dangerous forms of cancer.1 Many patients with oral cancer do not receive treatment until the disease has reached an advanced stage, resulting in a poor prognosis. Early screening and diagnosis of oral cancer can improve survival rates and reduce unnecessary costs.2 The development of oral cancer is a long process that passes through a “precancerous” period. Oral potentially malignant disorders (OPMDs), as defined by the World Health Organization in 2020, are any oral mucosal abnormalities associated with a statistically increased risk of developing oral cancer.3 These include oral leukoplakia, proliferative verrucous leukoplakia (PVL), erythroplakia, oral submucosal fibrosis, oral lichen planus (OLP), actinic keratosis, palatal changes in reverse smokers, discoidal lupus erythematosus, dyskeratosis congenital, oral lichenoid lesions, and oral chronic graft-versus-host disease. OPMDs is considered a general concept because all 11 diseases carry a risk of progression to oral cancer, although not all cases eventually become malignant.3 Therefore, it is crucial to identify lesions that are likely to undergo malignant transformation. Visual examination is the most common and intuitive method for this purpose. Additionally, special staining and fluorescence examination are also used to identify OPMDs.4–6 Pathological examination remains the gold standard for diagnosing OPMDs, but it has the drawbacks of invasiveness and operational difficulty.7 Furthermore pathologistsmust spend considerable time and effort analyzing images from pathological slides, and the influence of the examiners subjective judgment cannot be ignored.8
The concept of artificial intelligence (AI) was first proposed in the 1950s. AI, which originates from computer science, refers to a set of theories, methods, technologies, and application systems that simulate, extend, and enhance human intelligence. AI attempt to develop computer systems that imitate human work and thought processes. The strength of AI lies in its ability to learn and identify patterns and relationships from large, multidimensional, and multimodal data sets. However, due to limitations in computing power and data availability, the application of AI has faced many challenges.9 Machine learning is one of the derivative technologies of AI and is an essential condition for machines to achieve intelligence.10 Deep learning (DL) enables the processing of more complex data by increasing the number of hidden layers in artificial neural network algorithm. It represents a more advanced stage of machine learning. DL specializes in discovering complex structures in multi-dimensional data and extracting features, overcoming the data size limitations inherent in traditional ML.11 DL technologies include deep belief network convolutional neural network (CNN), and recurrent neural network. The advantage of CNN is that it can directly input data such as images, automatically segment and extract features from each part of the data, and integrate convolution layers for data processing, thus avoiding errors caused by manual input and minimizing susceptibility to interference. The number of image features that can be recognized by CNN has increased significantly. As a result, the introduction of CNN has brought substantial improvements to fields such as image processing and natural language processing, and holds great potential for medical applications requiring image recognition, such as medical imaging and histopathology.12 With the rapid development of AI-related technology, its application in the medical field is expected to become increasingly widespread. This review provides an overview of emerging DL techniques and their applications in the diagnosis and prognosis of OPMDs.
Application of DL in the auxiliary diagnosis and screening of OPMDs
There are various clinical examination methods for OPMDs, such as toluidine blue staining, autofluorescence, exfoliative cytology, chemiluminescence, and pathological examination. The results of these tests are typically reviewed by experienced doctors, which can make the diagnosis subjective. DL technology, represented by CNN algorithm, has unique advantages in image processing, which is helpful for the DL of pathological or imaging image classification. The internal structure of CNNs is achieved by providing data, mainly model parameters, in a convolutional and repetitive manner. The training process is repeated until incremental improvements in model detection ability allow the input image to be mapped to a specific label.13 With the assistance of CNN, the computer automatically identifies the best features to match the target image by predesigning the learned feature variables and directly classifies the image without relying on a large amount of data preprocessing or human operation and interference. AI technology has been widely used in clinical practice. Figure 1 demonstrates that it can assist doctors in the diagnosis of oral cancer and precancerous lesions bycompressing, enhancing, reducing, matching, describing, and recognizing various images.14
DL in clinical photographic images of OPMDs
At present, visual examination and palpation are the most routine methods for detecting oral cancer and OPMDs. These two examination methods are intuitive and convenient but require a high level of diagnostic ability and experience from the doctor. Some less developed regions may not have specialists available to diagnose OPMDs by conventional methods.
Algorithms such as DL-based CNN and DenseNet have been used to detect lesions in clinical images of the skin and larynx with results comparable to those of experts.15–17 However, identifying OPMDs from photographic images is much more difficult than classifying skin lesions because mucosal lesions are often hidden or masked in a complex background by overlapping teeth, tongue, and palate.18 Despite facing more challenges, some DL algorithms have been successfully applied to clinical photographic image detection and analysis of OPMDs. In general, DL models are divided into segmentation models, image classification models, and object detection models. Segmentation models are used to distinguish the lesion area from normal tissue, image classification models usually identify whether the lesion is cancerous, and object detection models diagnose the lesion as a more specific disease. For instance, Warin et al.19 used DenseNet-169, ResNet-101, SqueezeNet, and Swe-S as classification models and adopted Faster R-CNN, YOLOv5, RetinaNet, and CenterNet2 as detection models to analyze 980 images (365 oral squamous cell carcinoma [OSCC] images, 315 OPMDs images, and 300 benign lesions images). Compared with the results of oral and maxillofacial surgeons and general practitioners, the sensitivity and specificity of DenseNet-169 and ResNet-101 classification models were better than those of specialists and general practitioners. The sensitivity of R-CNN was between that of specialists and general practitioners, indicating that the CNN algorithm-based model has an expertlevel ability to distinguish oral cancer and OPMDs from benign lesions.19 Keser et al.20 developed a CNN model based on GoogleNet Inception V3 to identify photographic images of OLP. They found that this model was highly effective in distinguishing normal mucosa from OLP lesions. Tanriver et al.21 realized that excellent object detection models and classification models could be concatenated for identifying oral lesions. In their study, they proposed an end-to-end two-stage model combining the classification model EfficientNet-b4 with the detection model YOLOv5l to classify lesions into three categories: benign, OPMDs, and oral cancer. This is a low-cost OPMDs screening method that can automatically detect and classify various types of oral lesions in realtime.21
Ferrer-Sánchez et al.22 used a U-Net-based lesion segmentation model and a multi-task CNN classifier model to analyze 261 clinical photographs of oral leukoplakia to predict the risk of epithelial dysplasia and malignant transformation. One of the innovations of this study was the construction of two attention heatmaps of the images to interpret the predictions made by the model. The results showed that for predicting malignant transformation, the model achieved a sensitivity of 1 and specificity of 0.692, while for predicting high-risk dysplasia, the model achieved a specificity of 0.740 and sensitivity of 0.928. These two attention heatmaps, explaining the risk of malignant transformation and epithelial dysplasia, respectively, greatly increased the confidence in the prediction model.22
With the popularity of smartphones, mobile phones have become one of the most convenient tools for taking pictures. In a retrospective study by Lin et al.,23 they used the HRNet model based on the CNN algorithm to analyze mucosal lesion images taken using smartphones. After standardized training in photographic methods, Lin et al.23 used the HRNet model to analyze 688 images of mucosal lesions (251 recurrent aphthous ulcers, 231 low-risk OPMDs, 141 high-risk OPMDs, and 65 oral cancer) and 760 images of normal mucosa. The results showed that the sensitivity of the HRNet model was better than other classification models, such as VGG16, ResNet50, DenseNet169, and HRNET-W18, though it still misdiagnosed 1.9% of high-risk OPMDs as recurrent aphthous ulcer.23 The missed diagnosis of high-risk diseases is more harmful than the misdiagnosis of low-risk diseases in screening, as patients may lose the best opportunity for treatment. Despite the good results of the classification model in Lin’s study, the ability of the CNN model to analyze images taken by smartphones compared with specialists could not be determined. Without further optimization of the photographing method or classification model, smartphones cannot replace the role of digital single-lens reflex cameras in image acquisition.
Application to autofluorescence spectrum analysis
Autofluorescence is a safe and convenient method for screening OPMDs and early oral cancer. When normal tissues are exposed to blue light, they absorb part of the photon energy and emit lower-energy photons, a phenomenon known as autofluorescence. The fluorescence-producing molecules involved are mainly nicotinamide adenine dinucleotide, flavin adenine dinucleotide, and some elastin, which makes the image of normal tissue appear with green fluorescence. Abnormal tissues, due to changes in porphyrin metabolism and the breakdown of elastin, emit less fluorescence in areas at risk of malignant transformation, showing a lack of fluorescence, which appears black in the image.24 In a study by Morikawa, the evaluation of autofluorescence images was entirely subjective, showing high sensitivity (98.0%) and low specificity (43.2%) for detecting squamous cell carcinoma. Factors such as inflammation may interfere with the findings. Therefore, Morikawa concluded that a new evaluation method should be developed to obtain an objective evaluation.25
van Staveren et al.26 trained a DL model based on artificial neural network algorithms to analyze the autofluorescence spectrum of oral leukoplakia to determine its characteristics and grade of dysplasia. The results of the study showed that DL could effectively distinguish between the autofluorescence spectra of leukoplakia and normal tissue, as well as between homogeneous and heterogeneous leukoplakia.26 Although the interpretation of autofluorescence spectra by artificial neural network algorithms may not necessarily be the best method for evaluating autofluorescence detection, this study found that the edges of areas with fluorescence loss may not be distinguishable by the naked eye. Real-time intraoperative detection of the spectrum at each pixel can help the physician determine the extent of resection required for oral leukoplakia.
DL technology combined with exfoliative cytology
Exfoliative cytology is considered an effective method for mass screening of high-risk populations, and its role in cervical cancer screening has been well established.27 Previous research from our group has confirmed that exfoliative cytology hassignificant potential as an accurate and simple diagnostic method for clinically suspected oral precancer and oral cancer.28 In the traditional exfoliative cytology test, oral mucosal brush specimens are directly applied to glass slides for staining and observation. The analysis of cell morphology is a burdensome task that relies on experienced specialists.29,30
Sunny et al.31 used the Inception V3 model to analyze exfoliative cytology images of oral cancer and OPMDs, which was employed to automatically diagnose and risk-grade OPMDs. The test results of this model showed sensitivity and specificity of 73% and 100%, respectively. The study concluded that the tele-cytology platform combined with the CNN model improved accuracy by 30% compared to the traditional manual method.31
Exfoliative cytology is a qualitative rather than quantitative method for OPMD detection. Liu et al.32 developed the oral cancer risk index 2 (OCRI2) for the quantitative evaluation of leukoplakia. In this study, the peak random forest model was used to analyze the exfoliative cytology results of 68 patients with oral leukoplakia and calculate the corresponding OCRI2 values. The results showed that an OCRI2 value of 0.5 was the best threshold for predicting the malignant transformation of leukoplakia (OCRI2 values higher than 0.5 indicated high-risk leukoplakia, while values lower than 0.5 indicated low-risk leukoplakia). Using OCRI2 to quantitatively predict the risk of malignant transformation of leukoplakia based on exfoliative cytology can reduce the cost of patient follow-up.32 OLP may be severe on one side of the buccal mucosa and mild on the other side, but the mild side may still have a risk of malignant transformation.3,33,34 Despite this, it is clinically difficult to perform a pathological examination of the minor side mucosa, and noninvasive exfoliative cytology is acceptable. Therefore, the mild side mucosa of patients with OLP can be screened by DL-assisted exfoliative cytology when necessary to determine the subsequent treatment plan.
DL in pathological images of OPMDs
Many studies have shown that DL algorithms can assist pathologists in the diagnosis of oral malignant tumors.35 For instance, Aubreville et al.36 used the LeNet-5 model to identify confocal laser endomicroscopy images of OSCC, showing that the average accuracy was 88.3%, sensitivity was 86.6%, and specificity was 90%, which was even better than classifiers based on conventional AI algorithms.
Pathological examination is the gold standard for the diagnosis of OPMDs. AI-based analysis of histological images can minimize human interference and reduce the subjective influence of pathologists. Epithelial dysplasia is an important feature of malignant transformation in OPMDs, and the nuclei of these abnormal epithelial cells undergo varying degrees of change. Alshawwa et al.37 used multiple CNN models to analyze changes in nuclear entropy in tissue sections to identify oral leukoplakia and PVL. The Mask R-CNN model was used to segment the nucleus images and extract image features. The average accuracy of the model in segmenting the nuclei of leukoplakia, PVL, and SCC was 92.95%. The use of polynomial classifiers to distinguish leukoplakia and PVL also achieved good results, with average sensitivity of 95.83%, average specificity of 98.29%, and average accuracy of 97.05%, demonstrating that the classifier could distinguish between the two lesions.37,38 Idrees et al.38 combined whole slide imaging with computer-aided image analysis technology to construct an artificial neural network-multilayer perceptron to analyze and identify OLP. The principle of this study is that the total number of monocytes and granu locytes increases in OLP lesions. The results showed that the model could determine the critical point between OLP and other lichi-like diseases based on the number of inflammation cells, with a sensitivity of 100% and an accuracy of 94.62%.38 It is well known that AI mainly relies on the segmentation and identification of nuclei to analyze pathological images. Therefore, compared with the interpretation of nuclear atypia, the detection of OLP by inflammatory cell count is less reliable due to the difficulty in differentiating it from other chronic inflammatory diseases, and related studies are rarely reported.
DL for optical coherence tomography (OCT) images
OCT is a non-invasive and radiation-free real-time imaging method. The advantage of OCT is that it provides real-time images with a resolution comparable to that of pathological examination, which can be used to determine the edge of the tumor.39,40 The principle of OCT is similar to that of ultrasound examination, where electromagnetic waves of a certain wavelength pass through the tissue, and the optical properties of the tissue determine the optical path and depth of the light. Compared with normal tissues, tumors and precancerous lesions have a larger nuclear/plasma ratio, widened epithelial spikes, and a thickened basement membrane, leading to speckle-like images on OCT.41 The major drawback of OCT is that its evaluation depends on the operator’s professional knowledge. Since the imaging configuration differs from conventional radiological examinations, many specialists find it difficult to evaluate the image reports.
James et al.42 used a DL-based Support Vector Machine (SVM) to automatically diagnose 347 OCT images (151 normal oral mucosa, 121 OPMDs, and 75 malignant lesions). The model achieved a sensitivity of 93% and specificity of 74% for OPMDs, and a sensitivity of 95% and specificity of 76% for malignant tumors. The Inception-ResNet-v2-SVM model showed the highest sensitivity (83%) in differentiating mild dysplasia from moderate to severe dysplasia. These studies indicate that the accuracy of the SVM model in interpreting OCT images is comparable to that of pathological examination.42 Heidari et al.43 developed a CNN model to classify OCT images of OSCC and dysplasia in order to determine the boundary between normal and abnormal tissues in 3D images. Surprisingly, the sensitivity and specificity of the model were 100% and 70%, respectively, while the sensitivity and specificity of expert pathologists were 85% and 78%, respectively. Although the performance of the CNN model is close to that of pathologists, the model may misjudge uninvolved tissue as abnormal.43
Application of DL in predicting and monitoring the prognosis of OPMDs
With the advent of AI in medicine, DL techniques have become a common method for predicting disease development and outcomes based on informative data. Many studies have reported the use of DL algorithms to predict tumor development, including malignant transformation, lymph node metastasis, and prognosis.44,45 These algorithms learn from health data to provide automatic and exclusive prediction or classification of clinical outcomes without direct programming by the user. Many products based on this technology are used in precision medicine to support clinical decision-making and encourage individualized treatment choices for patients.46–48
For OPMDs, it is important to estimate the rate of malignant transformation of lesions when designing treatment. Unfortunately, the malignant transformation rate of many OPMDs, including oral leukoplakia, is highly variable. Clinicians should consider the location, size, color, and other characteristics of the lesion, as well as the grade of epithelial dysplasia, to analyze the prognosis. Most models used to predict the prognosis of oral cancer focus more on tumor metastasis, clinical outcomes, and treatment effects, while they often overlook the malignant transformation of OPMDs. Some models that claim to predict the rate of malignancy in OPMDs simply classify the predictions into high- and low-risk categories, rather than including real-time data as a dynamic variable of transformation probability over time. Obviously, the latter has greater clinical application value but is less explored in research.
Wang et al.49 developed two random forest classification models: the baseline model and the personalized model (model-P), to predict the malignant risk level of OPMDs. This study collected personal information (age, gender, lifestyle, and lesion status), non-invasive oral examinations (toluidine blue staining and autofluorescence), oral tissue biopsies, histopathological analysis, and treatment options. After comparing the performance of the baseline model, model-P, and clinical experts in predicting the risk of malignant transformation, the specificity of all three was comparable (about 90%), while the sensitivity of model-P was better than the other two (over 80%). The study also found that the scope of autofluorescence loss, toluidine blue staining score, and degree of lesion infiltration were important factors affecting carcinogenesis.49 Adeoye et al.50 compared three DL models (DeepSurv, time-dependent neural network Cox model, and DeepHit) with two conventional statistical models (random survival forest and Cox proportional hazards) to predict the risk probability of malignant transformation of OPMDs. The predictors entered into these models were gender, age, history of smoking and drinking, abstinence from bad habits, history of cancer, family history, hypertension, hyperlipidemia, diabetes, autoimmune disease, viral hepatitis, and the location and classification of the lesion. The DeepSurv algorithm was found to have the best discrimination performance when simulating the malignant transformation risk of oral leukoplakia and oral lichenoid lesions. However, the conventional random survival forest model outperformed other models in probability calibration.50
Difficulties in image analysis and AI algorithms
Automatic classification of medical images is one of the main applications of AI in the medical field. A recent review of AI/ML-based medical devices approved in the United States and Europe between 2015 and 2020 found 126 devices approved or CEmarked for radiological use in Europe and 129 devices in the United States, both representing more than half of the total reviewed.51 However, the identification and analysis of medical images by AI require the corresponding image database. Unfortunately, large databases of OPMDs and oral cancer have not been established for either clinical photographic images or pathological images, which limits the promotion of current AI-related technologies. Although Warin et al. found that DenseNet121, a newly developed deep CNN model, can reduce task overfitting caused by small databases, the ongoing challenges of database limitations still cannot be ignored.47,52,53 For this reason, Davenport et al.47 claim that humans are unlikely to see substantial changes in healthcare employment due to AI over the next 20 years. To address low-resolution images, Chen et al.54 proposed a joint framework called SRFBN+, which contains a novel transfer learning strategy and a deep super-resolution framework for generating high-resolution slice images from low-resolution slice images. The test results show that SRFBN+ performs well in generating super-resolution pathological examination images.54 Most of the DL-based models for detecting OPMDs rely on various clinical images, and the quality and standardization of these images present a major challenge for future AI applications.
The CNN algorithm has unique advantages in image processing, which is helpful for DL of pathological or photographic image classification. With the assistance of CNN, the computer automatically identifies the best features to match the target image by predesigning the learned feature variables and directly classifies the image without relying on a large amount of data preprocessing or human operation.14 However, CNN has two obvious drawbacks: first, training and testing the model takes a lot of time, and second, details may be lost due to glare when processing images. It is crucial to test how well the model works, and evaluation includes three aspects: statistical validity, clinical utility, and economic utility.48 Therefore, improving the time-consuming problem remains difficult. The loss of detail in resolution is related to the normalization method and the convolutional layer architecture of the CNN. Both the imaging equipment and methods affect the quality of the images, which in turn impacts the auxiliary diagnosis results of the AI model. This calls for the development of a set of image quality standards similar to those being developed for imaging. The problem of convolutional layer architecture can only be improved with the help of other DL algorithms, such as Fully Convolutional Networks.35
For the algorithm itself, it is better to propose a new algorithm than to optimize the one already in use, as the latter may lead to the conclusion that some subdomain algorithms are not improving.35 The development of new algorithms requires creating a multidisciplinary team that includes computer and social scientists, operations and research leaders, clinical stakeholders (physicians, caregivers, and patients), and experts in related disciplines. However, algorithm optimization must take various trade-offs into account due to the limitations of data and computing power.
Currently, some studies use manual cropping of the region of interest to preprocess images, while other studies use segmentation models to divide images into patches of a certain size to avoid errors in manual process.55 However, the limitation of this method is that the network cannot directly analyze the image as a whole and can only focus on a small part, which may also account for the loss of details. Wirtz et al.56 believe that image segmentation is limited by the amount of data and computing power. In the future, algorithms that can analyze the whole image without human intervention in the preprocessing stage should be developed, allowing for more accurate results per unit of computational cost.56
Discussion
In an era of increasing digitization of human health data, DL is expected to play a significant role in the development, validation, and implementation of decision support tools to facilitate precision medicine. In this review, we have demonstrated many promising applications of DL in various fields of OPMDs, including digital clinical photographic image analysis, autofluorescence spectroscopy analysis, exfoliated cell examination, pathological section analysis, and OCT image analysis for auxiliary diagnosis. Additionally, we discussed the performance of different DL models in predicting the prognosis of OPMDs (Table 1).19–23,26,31,32,37,38,42,43,49,50 Currently, most DL models for the auxiliary diagnosis of OPMDs appear to have more potential in screening, and the use of AI technology to interpret exfoliative cytology is one of the most convenient and promising screening methods. The promotion of these technologies could address some of the difficulties associated with the diagnosis of OPMDs in medically underserved areas, potentially reducing the workload of clinicians and pathologists and minimizing the interference of their subjectivity. However, DL models still face challenges such as insufficient image databases, low resolution of images, and limitations in the performance of the algorithms themselves. In the future, there may be technologies that diagnose OPMDs based on salivary markers, with corresponding DL models potentially offering more meaningful screening. Another emerging frontier is histologic inference of genomic features. As research progresses, the future of DL applications in the field of OPMDs will likely focus on multimodal learning to integrate medical images and omics data to identify biologically meaningful biomarkers.57 AI algorithms for predicting patient prognosis have long been applied in oncology,58–60 and even some AI models in commercial treatment planning systems, such as RapidPlan and Auto-Planning, can automatically plan and design treatment strategies.61,62 Unfortunately, few of these frontier studies have been applied to the field of OPMDs. In the future, integrated AI algorithms capable of analyzing mucosal lesions and automatically generating diagnostic reports and evaluating treatment options may be developed. Rapidly evolving DL technologies are expected to continue having a significant impact on the field of OPMDs in the near future. Both researchers and physicians need to be prepared for this revolutionary era.
Table 1Summary of DL models for OPMD diagnosis and prognosis
| Number | Author (Year, Country) | Purpose | Learning machine | Sample type | Sample number or information | Main outcome | References |
|---|
| 1 | Warin et al (2022, Thailand) | To use a novel deep convolutional neural network for early detection of oral cancer and oral lesions. | DenseNet-169, ResNet-101, SqueezeNet, Swe-S, Faster R-CNN, YOLOv5, RetinaNet, CenterNet2 | Clinical photographic image | 365 images of OSCC, 315 images of OPMDs, 300 images of benign lesions | The model based on CNN algorithm had expert level in distinguishing oral cancer and OPMDs from benign lesions. | 19 |
| 2 | Keser et al (2022, Turkey) | To develop a deep learning method to identify oral lichen planus lesions using photographic images. | GoogleNet Inception V3 | Clinical photographic image | 65 images of healthy oral mucosa, 72 images of the buccal mucosa of OLP | The deep learning model provided classification of all tested images of healthy and diseased mucosa with 100% success rate. | 20 |
| 3 | Tanriver et al (2021, Turkey) | The potential application of computer vision techniques in the field of oral cancer within the range of photographic images, and the promise of an automated system for the detection of OPMD was investigated. | EfficientNet-b4, YOLOv5l | Clinical photographic image | 162 images of OSCC, 248 images of OPMDs, 274 images of benign lesions | Preliminary results demonstrate the feasibility of the deep learning-based method for real-time automatic detection and classification of oral lesions. | 21 |
| 4 | Ferrer- Sanchez et al (2022, Spain) | To estimate the probability of malignancy in oral leukoplakia using deep learning. | U-Net based lesion segmentation model, multi-task CNN classifier model | Clinical photographic image | 261 images of OLK | The sensitivity and specificity of the model in predicting malignant lesions were 1 and 0.692, respectively. The specificity and sensitivity of predicting high-risk dysplasia were 0.740 and 0.928, respectively. | 22 |
| 5 | Lin et al (2021, China) | A DLD based smartphone image diagnosis method for automatic detection of oral diseases | HRNet-W18 | Clinical photographic image | 251 images of ulcer, 231 images of low-risk OPMDs, 141 images of high-risk OPMDs, 65 images of oral cancer, 60 images of normal mucosa | The sensitivity of HRNet model was better than that of VGG16, ResNet50, DenseNet169 and HRNET-W18 classification models. | 23 |
| 6 | van Staveren et al (2000, Netherlands) | A deep learning model based on artificial neural network algorithm was used to analyze the autofluorescence spectra of oral leukoplakia to determine the characteristics and grading of dysplasia. | Trained ANN model | Autofluorescence spectrum | 22 images of OLK, 6 images of normal mucosa | Abnormal tissues can be distinguished from normal tissues by neural networks with a sensitivity of 86% and a specificity of 100%. In addition, the classification of homogeneous or heterogeneous tissue performed reasonably well. | 26 |
| 7 | Sunny et al (2019, India) | The risk stratification model based on ANN was used for the early detection of OPMDs. | Inception V3 model | Exfoliative cytology image | 981 images | The sensitivity and specificity of the model were 73% and 100%, respectively. The remote cytology platform combined with CNN model improved the accuracy by 30% compared with the traditional manual method. | 31 |
| 8 | Liu et al (2017, China) | An oral cancer risk index using DNA index values was developed to quantitatively assess cancer risk in patients with oral leukoplakia. | Peak random forest model | Exfoliative cytology image | 18 normal, 28 OLK, 41 OSCC subjects | OCRI2 can distinguish between low-risk and high-risk OLK and may improve the cost-effectiveness of patients with OLK during clinical follow-up. | 32 |
| 9 | Alshawwa et al (2022, Saudi Arabia) | Tissue sections were analyzed using multiple CNN models to differentiate oral leukoplakia from proliferative verrucous leukoplakia. | Mask R-CNN, polynomial classifier | Pathological images | 568 OLK images, 45 PVL images, 58 SCC images | The average accuracy of Mask R-CNN model in the nuclear segmentation of leukoplakia, PVL and SCC was 92.95%. The average sensitivity of the polynomial classifier was 95.83%, the average specificity was 98.29%, and the average accuracy was 97.05%, indicating that the classifier could distinguish between the two lesions. | 37 |
| 10 | Idrees et al (2021, Australia) | A machine-learning artificial neural network was created to detect OLP by identifying and quantifying monocytes and granulocytes in inflammatory infiltrates in digitized hematoxylin and eosin microscope slides. | ANN-MLP | Pathological images | 1,606 images of OLP | The machine learning method was able to reliably detect the critical point between OLP and other lichi-like diseases based on the number of inflammatory cells and monocytes, with a sensitivity of 100% and an accuracy of 94.62%. Artificial intelligence has the potential to improve the accuracy of pathologists in the diagnosis of OLP. | 38 |
| 11 | James et al (2021, India) | A point-of-care optical coherence tomography device was used to detect potentially malignant and malignant lesions in the oral cavity. | Support Vector Machine (SVM), Inception-ResNet-v2-SVM model | OCT images | 75 images of malignant lesions, 121 images of OPMDs, 152 images of benign lesions | The sensitivity and specificity of the SVM model for OPMDs were 93% and 74%, and the sensitivity and specificity for malignant tumors were 95% and 76%. The Inception-ResNet-v2-SVM model had the highest sensitivity (83%) in differentiating mild dysplasia from moderate and severe dysplasia. This study shows that the accuracy of the SVM model in OCT image interpretation is close to that of pathological examination. | 42 |
| 12 | Heidari et al (2020, USA) | A CNN model was developed to classify OCT images of OSCC and dysplasia to determine the boundaries between normal and abnormal tissues in 3D images. | Trained CNN model | 3D OCT images | 1,000 images from 6 cases of oral cancer and 1 case of dysplasia | The sensitivity and specificity of the model were 100% and 70%, respectively. This approach has the potential to be used as a real-time analysis tool to assess surgical margins. | 43 |
| 13 | Wang et al (2020, China) | To predict the malignant risk level of OPMDs. | Baseline model (model-B), personalized model (model-P) | 266 patients with OPMDs | Personal information, non-invasive oral examination, oral tissue biopsy and histopathological analysis, treatment and follow-up | The specificity of the two models was comparable to that of experts (about 90%), while the sensitivity of Model-P was better than the other two models. | 49 |
| 14 | Adeoye et al (2021, Hong Kong) | To compare the ability of three deep learning models and two traditional statistical models in predicting the risk probability of malignant transformation of OPMDs. | DeepSurv, time-dependent neural network Cox model, DeepHit, random survival forest, Cox proportional hazards | 716 patients with a clinical diagnosis of oral leukoplakia, oral lichen planus, or oral lichenoid lesions | Gender, age, history of smoking, drinking, abstinence from bad living habits, history of cancer, family history, history of hypertension, history of hyperlipidemia, history of diabetes, history of autoimmune disease, history of viral hepatitis, lesion location and classification | DeepSurv and RSF are two excellent performing models based on the discrimination and calibration after internal validation, and DeepSurv is more stable than RSF when cross-validated. Overall, the time-to-event model successfully predicted malignant transformation of oral leukoplakia and oral lichenoid lesions, | 50 |
Conclusions
This review summarizes five image analysis methods based on deep learning neural networks for the diagnosis of OPMDs and the prediction of malignant risk. Furthermore, the current limitations and future development prospects of deep learning in OPMDs are evaluated. The combination of these emerging technologies and diagnostic methods will change the clinical diagnosis and treatment of OPMDs.
Declarations
Funding
This work was supported by grants from the National Natural Science Foundation of China (No.82270983, No.82470982).
Conflict of interest
The authors declare that they have no conflict of interest.
Authors’ contributions
Writing - original draft (XLL), writing - review & editing, supervision, funding acquisition (GZ). Both authors have approved the final version and publication of the manuscript.