A Novel Method for the Detection of Oral Cancers: Deep Learning

doi:10.21203/rs.3.rs-3123156/v1

Background:Oral cancers are among the most common cancers worldwide. Deep learning (DL) represents a technology that emulates human characteristics through programmed algorithms. In our study, oral cancer photographs were presented to dental professionals and artifical intelligence (AI) systems for diagnosis.

Materials and Methods: Verified photographs of oral lesions and healthy oral tissues were obtained from scientific publications, internet search engines, and personal archives. AI systems were trained using DL networks. The trained models were tested on a test set of photographs that were not used for training. The objective was to detect and categorize the lesions. Dental professionals were also asked to categorize the lesions based on the test set photographs. The performances of AI systems and dental professionals were compared.

Results:The survey included responses from 154 dentists, with an average score of 0.871 ± 0.110. Among dentists, periodontists had the highest success rate, with an area under the receiver operating characteristic curve (AUC) value of 0.937 ± 0.083. However, no significant differences were observed in the success rates among the dentists based on their specialty, years of professional experience, or previous experience with oral cancer monitoring (p <0.05). Based on the test conducted on 32 images, the most successful models were ResNet-101 and Inception v3, with an AUC value of 0.958.

Conclusion: The rapid advancements in DL are associated with numerous advantages in various fields, including healthcare, especially with regard to achieving accurate diagnoses. These benefits also extend to oral cancers. We found that data-driven AI systems can make a reliable and objective diagnosis of oral cancer.

Artificial intelligence deep learning

oral cancers

Oral cancers are among the most common cancers worldwide. The incidence of oral cancer continues to rise in numerous populations. The rapidly increasing incidence rate, delayed diagnosis, and inadequate treatment planning contribute to a worsened prognosis for oral cancer [1]. Early detection of oral cancer reduces the morbidity and mortality rates and improves the quality of life to prolong patient survival. Despite remarkable advancements in the field of medicine, persistent gaps in public knowledge regarding oral cancer, challenges in accessing healthcare in rural areas, and delayed diagnosis have led to an increasing incidence rate of oral cancer.

Despite the importance of comprehensive routine head and neck examination for the assessment of oral tissues, histopathological diagnosis is required to diagnose oral cancers. Therefore, auxiliary applications that aid in the detection of oral cancers are needed. Recent advancements in image analysis and diagnostic technologies and artificial intelligence (AI) have emerged as valuable tools in this context. These developments offer the ability to detect and classify lymph node involvement, a vital factor influecning the prognosis of oral cancers, and to differentiate between malignant lesions and lesions with malignant transformation potential [2]. The present study evaluated the usefulness of AI technologies for the diagnosis of oral lesions and classification of early oral cancers.

Oral cancers encompass malignant lesions that involve lips, buccal mucosa, hard palate, floor of the mouth, anterior two-third of the tongue, and gingiva surrounding the upper and lower teeth [3]. According to data from the International Agency for Research on Cancer, in 2020, oral cancer was diagnosed in 377,713 individuals worldwide, with 177,757 associated deaths [4]. Squamous cell carcinoma accounts for approximately 90% of all oral cancers [5]. While numerous factors influnence the prognosis of oral cancer, tumor stage is the most crucial factor [1].

Despite the advancements in diagnostic techniques and treatments, approximately one-third of oral cancers are diagnosed after metastases (stage 3 or 4), with an estimated 5-year survival rate of almost 50% [6].

Genetic predisposition and exposure to environmental carcinogens are important risk factors for oral cancer. In addition to malignant lesions, it is cruical to identify the malignant transformation potential of premalignant lesions [7]. The rate of malignant transformation of premalignant lesions varies widely, ranging from 0.13–85% [8]. Many lesions with malignant transformation potential are asymptomatic, leading to delays in the diagnosis, which negatively impacts the prognosis [1]. Comprehensive head and neck examination by healthcare professionals is essential to prevent such delays [9]. However, relying on clinical oral examination alone is inappropriate, as more than 30% of patients have undergone such examinations within 3 years prior to diagnosis [10]. Therefore, it is essential to develop auxiliary methods that can enhance the effectiveness of routine examinations and facilitate lesion detection.

Several techniques are used for the diagnosis of oral cancers, including oral exfoliative cytology, vital tissue staining, and light-based systems. Although exfoliative cytology is convenient and comfortable for patients, its reliability is unclear [11]. Toluidine blue demonstrates rapid cell division in inflammatory, regenerative, and neoplastic tissues; however, its sensitivity and specificity exhibit significant variations [12]. Light-based diagnostic systems evaluate suspicious oral mucosal tissues based on the absorption and reflection when exposed to light or energy. However, their routine use is limited due to low specificity rates, challenges in selecting the biopsy area, and high costs [13]. In recent years, auxiliary diagnostic methods have been integrated with AI for the early diagnosis of oral cancers [14].

The advancement of computer technology integrated with AI has revolutionized various fields, including medicine, mathematics, automotive, legal consulting, and security applications [15]. AI involves machines that exhibit human-like characteristics, primarily in cognitive abilities. Convolutional neural networks, a type of deep learning (DL) method, enable machines to learn effectively from images. DL relies on self-learning through the creation of multi-layered models that are trained on large datasets. Increasing the quantity of data can enhance the accuracy of classification or prediction [16, 17].

The protocol for this prospective, randomized study was approved by the Karamanoğlu Mehmetbey University Clinical Research Ethics Committee (No: 06-2023/08). Informed consent was obtained from the participating dental professionals.

Study design and setting

Oral lesions were categorized as benign, premalignant, or malignant. In total, 1,375 photographs were obtained from scientific publications, internet search engines, and personal archives. These photographs included confirmed images of lesions and healthy oral tissues. The photographs were cropped to focus on the oral region. Approximately 30% of the photographs from each group were randomly set aside for validation purposes. Eight photographs were randomly selected from each group for comparison (n = 32). As the effectivness of AI training is closely linked to the size of the input dataset, the remaining photographs were used for this purpose (Table 1).

Table 1

Distribution of the obtained photographs according to their intended use.
Class	Training	Validation	Test	Total
Benign	235	101	8	344
Premalign	243	104	8	355
Malign	359	154	8	521
Healthy	103	44	8	155
				1375

AI training was performed using various DL networks, including SquezeNet, GoogleNet, ResNet-50, EfficientNet, DarkNet-53, DarkNet-19, ShuffleNet, Xception, MobileNet v2, NasNet Mobile, DenseNet-201, ResNet-18, VGG16, VGG19, ResNet-101, and Inceptionv3, and MatLab software (R2023a; Mathworks, Natick, MA, USA). Each AI model required input photographs with specific resolutions. Therefore, the software resized the input photographs to the resolutions specified in Table 2. The training process involved 20 epochs, with a learning rate of 0.001 for the first 10 epochs and 0.0001 for the subsequent 10 epochs. A batch size of 16 and the sgd (stochastic gradient descent) optimization method were used. To evaluate the performance of the trained models in lesion detection and categorization, they were tested on unused test photographs that had been reserved for this purpose. Additionally, a questionnaire was developed using the photographs from the test set, and general dental practitioners and specialist dentists were requested to classify the lesions. Subsequently, the performance of AI was compared to that of dental professionals. To allow comparison, the true positive, true negative, false positive, and false negative rates were determined for each AI model and professional group. SPSS software (IBM Corp., Armonk, NY, USA) was used to calculate the sensitivity (ability to detect individuals with health problems) and specificity (ability to detect healthy individuals), which are essential parameters for diagnostic tests. The area under the receiver operating characteristic curve (AUC), another important parameter for diagnostic tests, was also calculated.

Table 2

AUC values of visual classification artificial intelligence models in detecting and classifying oral cancers.
CNN model	Resolution	AUC
ResNet-101	224x224	0,958
Inception v3		0,958
ResNet-50	224x224	0,932
DenseNet-201	224x224	0,93
MobileNet v2	224x224	0,93
GoogleNet	224x224	0,917
DarkNet-19	256x256	0,896
ShuffleNet	224x224	0,844
EfficientNet		0,839
VGG-16	224x224	0,831
NasNet Mobile	224x224	0,823
ResNet-18	224x224	0,813
Inception v2	299x299	0,807
DarkNet-53	256x256	0,802
VGG-19	224x224	0,771
SquezeNet	227x227	0,734
Xception	299x299	0,708

The AUC value, which ranges from 0 to 1, is a measure of the diagnostic method’s or individual’s success in making diagnoses. Table 2 presents the AUC values of DL-based AI models used for image classification. Based on the test conducted on 32 images, the most successful models were ResNet-101 and Inception v3, both of which achieved an AUC value of 0.958. Table 3 presents the average success of the 154 dental professionals in classifying lesions according to their areas of expertise. The average AUC value of the dentists was 0.871 ± 0.110, which was higher than the overall average. Among dentists, periodontists achieved the highest success, with an AUC value of 0.937 ± 0.083. Furthermore, no significant differences were observed in the success of dental professionals according to their field of specialization, years of professional experience, or previous experience with oral cancer surveillance (p < 0.05).

Table 3

AUC values of dental professionals in detecting and classifying oral cancers.
	n	Mean AUC	Std. Dev.	Std. Err.	95% Confidence Interval for Mean		Minimum	Maximum
	n	Mean AUC	Std. Dev.	Std. Err.	Lower Bound	Upper Bound	Minimum	Maximum
General Dentistry	82	0,859	0,108	0,012	0,835	0,882	0,512	1,000
Oral, Dental and Maxillofacial Surgery	25	0,903	0,127	0,025	0,851	0,956	0,604	1,000
Endodontics	5	0,767	0,076	0,034	0,673	0,861	0,668	0,852
Orthodontics	7	0,868	0,120	0,045	0,757	0,979	0,664	0,979
Pediatric dentistry	11	0,905	0,060	0,018	0,864	0,945	0,818	1,000
Periodontology	7	0,937	0,083	0,031	0,861	1,014	0,761	1,000
Prosthetics	11	0,844	0,115	0,035	0,767	0,922	0,644	1,000
Restorative Dentistry	6	0,916	0,073	0,030	0,840	0,993	0,815	1,000
Total	154	0,871	0,110	0,009	0,854	0,889	0,512	1,000

AI technologies can detect subtle changes that may not be detected by the human eye. By integrating data from diverse sources, they can assist in clinical diagnosis. The application of AI systems for the diagnosis of oral lesions has been employed for cancer screening, classification of suspicious mucosal changes, tissue diagnosis, evaluation of lymph node involvement, and gene expression analysis [18]. Logistic regression, support vector machines, decision trees, and Bayesian networks are among the most commonly used algorithms in oral cancer studies. In a systematic review of the effectiveness of machine learning-based algorithms in oral cancer, the accuracy rates were 89–97% for malignant transformation, 78–91% for cervical lymph node metastasis, 64–100% for treatment response, and 71–99% for prognosis [8]. In our study, AI achieved a success rate of 70–95% for diagnosing lesions.

A significant advantage of AI applications in cancer screening is the elimination of user fatigue and the ability to detect rapidly minor changes in a single pixel, which leads to improved workflow efficiency and increases the chances of early diagnosis [18]. By comparison, when presented with the images, dentists took longer to make a diagnosis than AI. A case-control study compared expert physicians with AI models based on baseline data, including sociodemographic factors, smoking status, and alcohol consumption, or genomic data of 84 oral cancer patients and 87 healthy individuals. Both models outperformed expert physicians in assessing the risk of oral cancer [19]. Similarly, Wang et al. [20] developed an algorithm that was aimed at assessing the risk of malignant transformation using data from 266 patients with suspicious oral lesions. The algorithm effectively predicted the risk of oral cancer. Dentists play cruical roles in the detection and diagnosis of oral cancer as they are involved in oral cavity examinations. To ensure the effectiveness of regular oral cavity examinations, dentists must possess sufficient knowledge regarding the various clinical presentations and risk factors of malignant and premalignant lesions and oral cancers [21, 22].

A previous study found no differences in the level of knowledge about oral cancer among experienced and inexperienced physicians [23]. Similarly, in our study, the experience of dentists did not have a statistically significant impact on their diagnostic accuracy.

Integration of AI-supported apps into mobile phones can effectively facilitate oral cancer screening in rural areas with limited access to healthcare services [24]. A previous study evaluated algorithms for the diagnosis and classification of oral cancers using intraoral photographs, and found that algorithms outperformed general practitioners and exhibited a performance similar to that of specialist physicians [25]. In our study, periodontists were the most successful group of dental professionals in diagnosing oral lesions. AI achieved similar levels of success to that of expert dentists. Based on a test conducted on 32 images, the most successful models were ResNet-101 and Inception v3, with AUC values of 0.958 for both. Therefore, DL models were more successful than expert dentists for the diagnosis of oral cancer.

Obtaining high-quality intraoral photographs can be challenging due to the anatomical structures surrounding the oral cavity, including lips, teeth, and buccal mucosa. In our study, we encountered difficulties associated with obtaining intraoral photographs. When the photographs used to train DL systems fail to capture accurately the diversity and heterogeneity of oral lesions, the algorithm’s ability to make precise predictions for such lesions also becomes compromised [18]. Song et al. [26] examined saliva samples collected from 124 healthy individuals, 124 individuals with premalignant lesions, and 125 individuals with oral cancer. The samples were analyzed using conductive polymer spray ionization mass spectrometry, to identify metabolites and distinguish metabolic pathways that undergo changes during malignant transformation. The machine learning algorithm achieved an accuracy rate of 86.7% in differentiating oral cancer and premalignant lesions from healthy mucosa, based on spectrometer measurements. In our study, the accuracy of AI based on intraoral photographs ranged from 70–95%.

Numerous studies have demonstrated that AI-based systems can be used as auxiliary diagnostic tools for oral cancer lesions. However, most studies used a retrospective approach and small sample size, which hampers the reproducibility of the results obtained from the available dataset [27]. Similarly, our study also included a small dataset.

Advancements in AI-based systems can promote their application for the diagnosis of oral cancer lesions, as well as in multiple other fields. Studies have shown that data-driven AI-based systems for the diagnosis of oral cancer are reliable, objective and, in some cases, more successful than humans. However, these systems are in the early or developmental stages, and their potential use for cancer diagnosis requires reproducibility. Given the challenges associated with early detection of oral cancer, AI systems can serve as a promising auxiliary method for the initial evaluation of oral lesions, facilitating timely diagnosis and treatment. These systems hold great potential for benefiting high-risk populations residing in rural or remote areas with limited access to healthcare services. Furthermore, it is crucial to emphasize the importance of establishing archives.

Ethics approval and consent to participate

All experimental protocols were specified in the ethical approval form and approval was obtained from the Ethics Committee of Karamanoğlu Mehmetbey University Faculty of Medicine for the study (Approval numbered: 06-2023/08). Informed consent was obtained from all volunteers and/or legal guardian(s). The authors verifies that all methods have been performed in accordance with relevant guidelines and regulations.

Consent for publication

Not applicable.

Availability of data and materials

The datasets used and analysed during the current study are available from the corresponding

author on reasonable request.

Competing interests

The authors declared that there are no possible conflicts of interest.

Funding

The authors paid for the materials and analysis fees used in the study.

Authors contributions

All support for the current article (e.g., funding, provision of study materials, medical writing, manuscript processing fees, etc.) has been covered by the authors.

Acknowledgements

Not applicable.

Authors information

Department of Oral and Maxillofacial Surgery Dentistry, Faculty of Ahmet Keleşoğlu Dentistry, Karamanoğlu Mehmetbey University, Yunus Emre Campus, Karaman/TURKIYE ORCID: 0000-0003-1920-4655

Elif Esra Özmen

Department of Restorative Dentistry, Faculty of Ahmet Keleşoğlu Dentistry, Karamanoğlu

Mehmetbey University, Yunus Emre Campus, Karaman/TURKIYE ORCID: 0000-0002-0840-7126

Türkay Kölüş

Department of Oral and Maxillofacial Surgery Dentistry, Faculty of Ahmet Keleşoğlu Dentistry, Karamanoğlu Mehmetbey University, Yunus Emre Campus, Karaman/TURKIYE ORCID:0000-0003-3112-8528

Veysel İçen

-Alabı RO, et al. Deep machine learning for oral cancer: from precise diagnosis to precision medicine. Front Oral Health. 2022;2:97.
-Özer NE, Karaca B, Güneri P. Artificial Intelligence Applications for the Early Diagnosis of Oral Cancer. J of Ege University Faculty of Dentistry.2022;19–26.
Van der Waal I, de Bree R, Brakenhoff R, Coebegh JW. Early diagnosis in primary oral cancer: Is it possible? Med Oral Patol Oral Cir Bucal. 2011;16:300–5.
- International Agency for Reesearch on Cancer 2020. globocan 2020. Availableonline: https://gco.iarc.fr/today/onlineanalysis-table? (accessed on 08 September 2022).
- Speight PM, Farthing PM. The pathology of oral cancer. Br Dent J. 2018;225:841–7.
- Chen SW, Zhang Q, Guo ZM, Chen WK, Liu WW, Chen YF, et al. Trends in clinical features and survival of oral cavity cancer: fifty years of experience with 3,362 consecutive cases from a single institution. Cancer Manag Res. 2018;10:4523–35.
- Warnakulasuriya S, Johnson NW, van der Waal I. Nomenclature and classification of potentially malignant disorders of the oral mucosa. J Oral Pathol Med. 2007;36:575–80.
- Adeoye J, Koohi-Moghadam M, Lo AWI, Tsang RK, Chow VLY, Zheng LW, et al. Deep learning predicts the malignant-transformation-free survival of oral potentially malignant disorders. Cancers (Basel). 2021;13(23):6054.
-Güneri P, Epstein JB. Late stage diagnosis of oral cancer: components and possible solutions. Oral Oncol. 2014;50:1131–6.
- Epstein JB, Güneri P, Boyacioglu H, Abt E. The limitations of the clinical oral examination in detecting dysplastic oral lesions and oral squamous cell carcinoma. J Am Dent Assoc. 2012;143:1332–42.
- Epstein JB, Feldman R, Dolor RJ, Porter SR. The utility of tolonium chloride rinse in the diagnosis of recurrent or second primary cancers in patients with prior upper aerodigestive tract cancer. Head Neck. 2003;25:911–21.
- Piazza C, Bon FD, Peretti G, Nicolai P. Biologic endoscopy’: optimization of upper aerodigestive tract cancer evaluation. Curr Opin Otolaryngol Head Neck Surg. 2011;19:67–76.
- Ilhan B, Guneri P, Wilder-Smith P. The contribution of artificial intelligence to reducing the diagnostic delay in oral cancer. Oral Oncol. 2021;116:105254.
- Ahmad Z, Rahim S, Zubair M, Abdul-Ghafar J. Artificial intelligence (AI) in medicine, current applications and future role with special emphasis on its potential and promise in pathology: present and future impact, obstacles including costs and acceptance among pathologists, practical and philosophical considerations. A comprehensive review. Diagn Pathol. 2021;16:24.
- Kaul, Vivek S, Enslin. Gross. History of artificial intelligence in medicine. Gastrointest Endosc. 2020;924:807–12.
-, Wu, Meiyin, Chen L. Image recognition based on deep learning. Chinese automation congress (CAC). IEEE; 2015.
- Xu B, Wang N, Chen T, Li M. Empirical evaluation of rectified activations in convolutional network. arXiv 2015; https://arxiv.org/abs/1505.00853.
- Tseng YJ, Wang HY, Lin TW, Lu JJ, Hsieh CH, Liao CT. Development of a machine learning model for survival risk stratification of patients with advanced oral cancer. JAMA Netw Open 2020; 3; e2011768.
- Rosma MD, Sameemii AK, Basir A, Mazlipahiv IS, Norzaidi MD. The use of artificial intelligence to identify people at risk of oral cancer: empirical evidence in Malaysian University. Int J Sci Res Edu. 2010;3:10–20.
- Wang X, Yang J, Wei C, Zhou G, Wu L, Gao Q, et al. A personalized computational model predicts cancer risk level of oral potentially malignant disorders and its web application for promotion of non-invasive screening. J Oral Pathol Med. 2020;49:417–26.
-Leao J, Goes P, Sobrinho C, Porter S. Knowledge and clinical expertise regarding oral cancer among Brazilian dentists. Int J Oral Maxillofac Surg. 2005;34(4):436–9.
- de Almeida Geraldino R, Rezende LVML, da-Silva CQ, Almeida JCF. Remote diagnosis of traumatic dental injuries using digital photographs captured via a mobile phone. Dent Traumatol. 2017;33:350–7.
Warin K, et al. AI-based analysis of oral lesions using novel deep convolutional neural networks for early detection of oral cancer. PLoS ONE. 2022;17:8.
-Kim JS, Kim BG, Hwang SH. Efficacy of artificial intelligence-assisted discrimination of oral cancerous lesions from normal mucosa based on the oral mucosal image: a systematic review and metaanalysis. Cancers (Basel). 2022;14:3499.
-Song X, Yang X, Narayanan R, Shankar V, Ethiraj S, Wang X, et al. Oral squamous cell carcinoma diagnosed from saliva metabolic profiling. Proc Natl Acad Sci USA. 2020;117:16167–73.
-Briganti G, Le Moine O. Artificial intelligence in medicine: Today and tomorrow. Front Med. 2020;7:27.

No competing interests reported.

A Novel Method for the Detection of Oral Cancers: Deep Learning

Status:

Version 1

Abstract

Background

Materials and Methods

Study design and setting

Results

Discussion

Conclusions

Declarations

References

Additional Declarations

Status:

Version 1