“Enhancing early detection of oral cancer: a comparative study of artificial intelligence models and clinical specialist in lesion classification” Authors: Shima Soleimani Sardou, Mohammad Mehdi Ghaemi, Fatemeh Sadat Rezvaninejad, Abolfazl Seyrfar, Arash Shahravan, Nader Navabi, Raziyehsadat Rezvaninejad Published: 2025 ABSTRACT AbstractBackgroundOral cancer remains a major global health issue, with timely diagnosis being essential due to its varied clinical presentation. This study explores how artificial intelligence (AI) can support early detection by analyzing intraoral photographs.MethodsA cross-sectional analysis was performed using 518 intraoral clinical images collected from the Department of Oral Medicine, Kerman Faculty of Dentistry, between 2009 and 2023. The dataset comprised 104 images of malignant lesions and 414 of benign or normal tissue, all confirmed by a specialist in oral pathology. Three pretrained deep learning models, DenseNet-121, EfficientNet-B0, and ResNet-50, were evaluated for their ability to classify lesions as malignant or benign. The data were split into training (80%) and testing (20%) sets, with preprocessing completed before analysis.ResultsAmong the models, DenseNet-121 demonstrated superior performance, achieving 91% accuracy, 75% sensitivity, 98% specificity, 75% positive predictive value, 96% negative predictive value, an F1 score of 84%, and an area under the curve of 90%. These results exceeded the diagnostic accuracy of an experienced oral specialist.ConclusionAI-based analysis of clinical images can significantly improve early oral cancer detection and should be integrated into clinical workflows to enhance diagnostic precision.Keywords: Oral cancer, Oral lesions, Early diagnosis, Artificial intelligence, Deep learning, Machine learning Introduction Oral cancer is a significant global health concern, with an estimated 377,713 new cases and 177,757 deaths annually worldwide [1]. Early detection of oral cancer significantly boosts the likelihood of successfully treating related lesions [2]. However, over half of oral and pharyngeal cancers are identified at advanced stages, which drastically lowers life expectancy, with a five-year survival rate of just 52% in the United States [3–6]. In contrast, early diagnosis can dramatically improve survival rates, potentially raising them to around 84% when the cancer is caught in its initial stages [7, 8]. These statistics highlight the difficulty of achieving early diagnosis of this disease due to the wide variation in symptoms among individuals and the challenge of distinguishing between malignant, pre-malignant, benign lesions, and normal features in the oral cavity based solely on clinical examinations, which remain the primary screening method. The limited effectiveness of current diagnostic approaches underscores the need for complementary methods to facilitate faster and more accurate diagnosis [9–11]. Artificial intelligence (AI), through machine and deep learning, enables computers to mimic human reasoning and has recently achieved remarkable accuracy in oral and dental diagnostics [12–14]. This technology has been utilized across all fields, including medical sciences, as well as in dentistry [15]. One of the most important impacts of artificial intelligence in dentistry is the early diagnosis and treatment of oral cancer [10]. AI addresses this challenge by enabling earlier and more accurate detection of potentially malignant lesions. By analyzing images from intraoral scanners and radiographs, AI algorithms can identify subtle changes that may indicate cancerous growth, often before they are visible to the naked eye [4]. These advancements facilitate timely interventions, which are crucial for improving survival rates [16]. In addition, AI tools assist in the evaluation of biopsy results by analyzing histopathological images with a high level of accuracy, complementing the expertise of pathologists. This leads to more precise diagnoses [17]. It is important to emphasize that artificial intelligence acts as an auxiliary diagnostic tool, assisting clinicians in detecting lesions rather than replacing their expertise [10]. Machine learning (ML) and deep learning (DL) are subsets of artificial intelligence [18]. Machine learning can be categorized into supervised learning, semi-supervised learning, and unsupervised learning [19]. Supervised learning uses labeled datasets for training, where these labeled datasets act as the “supervisor” for the algorithm [19]. The algorithm learns from the labeled input, extracts, and identifies common features from the labeled data to make predictions on unlabeled input [20]. In contrast, unsupervised learning works independently to find different features in unlabeled data [21]. Semi-supervised learning falls between these two, utilizing a small amount of labeled data along with a large amount of unlabeled data during training [22]. Based on the explanations provided, it can be concluded that machine learning can be widely used for predicting and diagnosing various types of cancers, including oral cancer, through labeled images and data [12]. Using this technology, lesions related to the oral cavity can be classified accurately and efficiently, similar to or better than a specialist. This allows for the differentiation of cancerous and oral potentially malignant disorders (OPMD) lesions from benign lesions and normal variations, ultimately contributing to the early diagnosis of oral cancer and subsequently improving the survival rates and quality of life for affected individuals [3, 12]. Considering the high prevalence of oral cancer in Iran and worldwide, as well as its significant impact on reducing the quality of life and lifespan of affected individuals [23, 24], and recognizing that early diagnosis is a crucial factor in improving the survival rates of these patients [25], we aim to conduct a study titled ‘Application of Deep Learning for the Early Diagnosis of Malignant Oral Lesions Using Intraoral Photographs.’ This study seeks to utilize new AI models and enhance the accuracy of previous models to help address the challenges faced by patients with oral malignancies. Materials and methods Data description This retrospective study received ethical approval from the Ethics Committee of Kerman University of Medical Sciences under the code IR.KMU.REC.1402.238. The data consisted exclusively of intraoral clinical photographs collected from the Department of Oral Medicine at Kerman Faculty of Dentistry. No radiographic images were included in this analysis. Inclusion criteria Photos in the folders of the Oral Medicine Department, Photos in the folders of the Oral Medicine Department, Photos related to lesions of the oral cavity, Photos related to lesions of the oral cavity, photos in JPEG or JPG format, or those convertible to these formats. photos in JPEG or JPG format, or those convertible to these formats. Exclusion criteria Low quality and unclear photos. Low quality and unclear photos. Photographs of areas other than the oral cavity, Photographs of areas other than the oral cavity, Photos whose pathology report shows evidence of epithelial dysplasia. Photos whose pathology report shows evidence of epithelial dysplasia. Data collection The study samples consist of two specific groups: Intraoral photographs of all patients diagnosed with oral and maxillofacial cancers. Intraoral photographs of all patients diagnosed with oral and maxillofacial cancers. Intraoral photographs of anatomical landmarks or lesions that were not diagnosed as cancerous. Intraoral photographs of anatomical landmarks or lesions that were not diagnosed as cancerous. The definitive diagnosis of the lesions was confirmed by oral pathology specialist through biopsy. All images were sourced from patients who were referred to the Department of Oral Medicine at Kerman Faculty of Dentistry between 2009 and 2023. As a routine procedure, patients’ consent was obtained before using images, with guarantees provided to maintain their privacy. A total of 518 clinical images of various oral lesions were utilized for this study (Fig. 1). The dataset was categorized into two groups: [1] “malignant” lesions, which included 104 images (20%) of squamous cell carcinoma (SCC) or other oral cavity malignancies, and [2] benign or normal lesions, comprising 414 images (80%). The clinic encompassed a range of conditions, but images related to dysplasia were excluded from our study because dysplasia, while an important indicator of potential malignancy, is often not directly considered in AI-based malignancy diagnoses due to its complex and variable nature [26]. Dysplasia refers to abnormal cell growth, which may or may not progress to cancer, making it harder for AI models and specialist to categorize consistently [27]. AI systems typically focus on more definitive markers, such as tissue invasiveness or clear malignancy signs, which are more easily identified in medical imaging or pathology reports [28]. Additionally, models often require clear, actionable data to ensure accuracy in predictions [28] The definitive diagnosis of all lesions was confirmed by an oral pathology specialist through microscopic (histopathological) examination, which was considered the gold standard for comparison with AI-based classification. Example of the benign and malignant images from the dataset: A both oral pathologist and oral medicine specialist diagnosed as Malignant, B Pathologist diagnosed as Malignant, but oral medicine specialist diagnosed as Benign Experiment Clinical images were collected by the authors from 223 patients over a span of 12 years using a Canon EOS 700 D digital camera. Consequently, images with varying sizes, zoom levels, angles, lighting conditions, and orientations were standardized and resized to 224 × 224 × 3, then converted to JPG format. The dataset was then randomly divided into two groups using random sampling in Python: [1] the training set, which contained 80% (n = 414 images), and [2] the test set, consisting of 20% (n = 104 images). Before model training, each image was visually inspected and cropped to focus on the lesion area according to clinical relevance. Cropping was performed manually by an oral medicine specialist to isolate the lesion region while excluding surrounding normal tissue. All background elements such as lips, skin, retractors, and gloves were removed to avoid bias and ensure that only mucosal surfaces were analyzed. Additionally, brightness and contrast were adjusted to achieve consistent illumination across images, and all files were normalized and converted to JPG format for uniform model input. A commonly used ratio in train-test split in ml modeling is 80:20, and it comes from the famous Pareto Principle (also known as the 80/20 rule), which suggests that in many situations, 80% of the results come from 20% of the inputs; this principle is often used as a heuristic for data splitting, with the larger portion allocated to training data to ensure sufficient learning, while still having a representative testing set. Image classification For this study, three Convolutional Neural Network (CNN)-based image classification networks, the DenseNet-121, EfficientNet-B0, and ResNet-50 transfer models were utilized to develop the binary-class image classification model [29–31]. The image classification experiments were conducted on Google Colab (Google Inc., CA, USA) utilizing an Nvidia Tesla T4 GPU, with driver version 535.104.05 and CUDA 12.1 (Nvidia Corporation, CA, USA). Prior to training, all images were preprocessed using Python 3.11 and open-source library scikit-learn 1.3.0 for stratified shuffled split [32], and resized to 224 × 224 pixels to match the input requirements of the neural networks. DenseNet-121 DenseNet-121, introduced by Huang et al. [33], is a deep convolutional neural network that introduces a distinctive connectivity pattern, where each layer is directly connected to every other layer ahead of it. Instead of the conventional sequential design—where each layer feeds its output only to the next—DenseNet-121 establishes shortcut connections from each layer to all subsequent layers in a feedforward fashion. This results in a total of 𝐿(𝐿+1)/2 direct connections for a network with 𝐿 layers. In practice, this means that the input to any given layer consists of the concatenated feature maps of all preceding layers. Similarly, the feature maps it generates are passed on to all following layers. This dense connectivity encourages extensive feature reuse across the network, resulting in highly parameter-efficient and compact architectures. Such properties make DenseNet-121 particularly advantageous for tasks requiring strong feature extraction, such as lesion classification in clinical images [33]. ResNet-50 ResNet-50, developed by He et al. [34], introduced the concept of residual learning, which addresses the degradation problem observed in very deep networks. As networks get deeper, simply stacking more layers often leads to worse performance due to vanishing gradients or overfitting. ResNet, short for Residual Network, mitigates this issue by introducing Skip Connections that allow gradients to flow more directly through the network during backpropagation. In ResNet-50, which contains 50 layers, the building block is the residual block, consisting of a series of convolutional layers with a skip connection that bypasses one or more layers. This design allows the network to learn only the residual function—the difference between the input and the desired output—rather than the full transformation, making optimization easier [34]. This innovation enables the training of much deeper networks by addressing the vanishing gradient problem, allowing ResNet-50 to maintain high accuracy even with increased depth. Its robust architecture has made it a widely used baseline in image classification challenges, including cancer detection [35]. EfficientNet-B0 EfficientNet-B0, proposed by Tan and Le [36], adopts a compound scaling method that uniformly balances network depth, width, and resolution. Unlike traditional scaling approaches, EfficientNet’s method leads to a family of models that are not only highly accurate but also computationally efficient. At the core of EfficientNet-B0 is the Mobile Inverted Bottleneck Convolution (MBConv), a type of building block adapted from MobileNetV2. Each MBConv block includes depthwise separable convolutions and squeeze-and-excitation (SE) optimization to improve feature recalibration. The major innovation of EfficientNet-B0 lies in the compound scaling strategy: instead of arbitrarily increasing depth (more layers), width (more channels), or resolution (larger input size) in isolation, EfficientNet scales all three dimensions together using fixed scaling coefficients, achieving a better balance between performance and resource usage [36, 37]. This makes it a strong candidate for applications like mobile-based or low-resource medical diagnostics. Three abovementioned pre-trained convolutional neural network models were adapted to output a single-dimensional vector suitable for binary classification tasks. The models used the Rectified Linear Unit (ReLU) as the activation function. The training process was configured with a maximum of 100 epochs, a batch size of 32, and a learning rate set to 0.001. Transfer learning is a widely employed technique in medical applications to address overfitting, particularly when dealing with small datasets that are susceptible to learning noise. In the training set, the number of clinical images for benign oral conditions was nearly four times greater than that for cancer, which can introduce bias and adversely affect the prediction performance of convolutional neural networks (CNNs), an issue known as “class imbalance” [38]. To mitigate the effects of class imbalance, we implemented a Stratified Shuffled Split to ensure both classes were equally represented in the training and test datasets [39]. Before entering the processing phase, labeled images (normal or malignant) were prepared by adjusting their size, format, color according to the analysis algorithm. Subsequently, the data were divided into two groups: training and testing, with a specific ratio of 80:20. The most effective machine/deep learning algorithms were selected based on methods from similar studies. The chosen algorithms were trained using the training data to differentiate between normal and malignant lesions, while the test data were used to evaluate the accuracy of the model. The evaluation focused on several parameters, including accuracy, sensitivity, specificity, positive and negative predictive values, F1 Score, and AUC and finally compared with the diagnoses made by an Oral Medicine specialist [40]. The performance of the model was assessed using standard classification metrics, defined as follows: Precision = TP/(TP + FP) Recall (Sensitivity) = TP/(TP + FN) Specificity = TN/(TN + FP) F1 Score = 2 × (Precision × Recall)/(Precision + Recall) Where: True Positive (TP): Cases correctly identified as malignant False Positive (FP): Non-malignant cases incorrectly predicted as malignant True Negative (TN): Non-malignant cases correctly identified False Negative (FN): Malignant cases incorrectly predicted as non-malignant Results Three algorithms DenseNet-121, EfficientNet-B0, and ResNet-50 were evaluated and compared with each other as well as with the diagnoses made by Oral Medicine specialist Table 1. The performance metrics used to evaluate the image classification models are precision (positive predictive value (PPV)), Negative predictive value (NPV), recall (sensitivity), specificity, F1 score, and area under the receiver operating characteristics (ROC) curve (AUC). Table 1Performance of different models on the test set (unseen data)PrecisionRecall (Sensitivity)SpecificityF1 ScoreAUCDenseNet-1210.750.760.980.840.90EfficientNet-B00.730.780.960.830.88ResNet-500.780.780.950.800.85Oral Specialist0.710.720.850.720.80 Performance of different models on the test set (unseen data) When employing the DenseNet-121, EfficientNet-B0, and ResNet-50 models, the results on test set (unseen data) were as follows: accuracy was 91%, 90%, and 90%, respectively; positive predictive value was 75%, 72%, and 78%; negative predictive value was 96%, 89%, and 89%; F1 score was 84%, 83%, and 79%; sensitivity was 75%, 78%, and 78%; and specificity was 98%, 95%, and 95%. The area under the curve (AUC) values for DenseNet-121 (0.90), EfficientNet-B0 (0.88), and ResNet-50 (0.85) all outperformed the diagnostic AUC of the Oral Medicine specialist (0.80), indicating superior discriminative ability on the unseen test set. The receiver operating characteristics (ROC) curves and AUC for each model are shown in Fig. 2. Ultimately, the results indicated that the DenseNet-121 transfer model exhibited superior accuracy across most parameters, aligning with findings from previous studies. The performance metrics presented in this section were evaluated on the test dataset comprising unseen samples, ensuring an unbiased assessment of the model’s generalizability and real-world applicability. Receiver Operating Characteristics (ROC) curves: a ResNet-50 model, b Efficient-B0 model, and (c) DenseNet-121 model Discussion Early detection of oral cancer is crucial for minimizing complications and reducing mortality rates [40]. Although the oral cavity is readily accessible for examination, oral cancers are still often diagnosed at advanced stages [41]. Utilizing artificial intelligence to identify suspicious changes in the oral mucosa presents a practical, cost-effective, and universal approach for the early diagnosis of oral squamous cell carcinoma (SCC) [42, 43]. This study investigates and compares the accuracy of three transfer deep learning algorithms: DenseNet-121, EfficientNet-B0, and ResNet-50 in identifying malignant and normal oral lesions using a dataset of confirmed clinical images. The results are also compared with the accuracy of an Oral Medicine specialist. Our findings indicate that the proposed convolutional neural network (CNN) can detect oral cancer lesions based on standard color images with high levels of accuracy (91%), sensitivity (75%), and specificity (96%), which is comparable to, and even surpasses, human performance (81%). According to our study, the diagnostic accuracy of a specialist in identifying suspicious oral lesions is estimated at 81%, while the accuracy reported for a general dentist in another study is 75% [44]. Similar studies employing CNN deep learning to analyze images in other cancer types, such as skin cancer, have reported comparable results, with sensitivity estimates ranging from 83% to 97% and specificity between 43% and 82% [45–47]. Despite these promising results, it is essential that AI-based diagnostic tools be used under the supervision of specialists, ensuring that algorithmic predictions are interpreted alongside professional clinical judgment. In machine learning, model accuracy depends on factors like data quality and quantity, where high-quality, ample data helps the model learn patterns better [48]. The choice of relevant features and the right model type also plays a critical role, as irrelevant features or unsuitable models can confuse or limit the model’s effectiveness [49]. Proper tuning of hyperparameters (e.g., learning rate) and thorough data preprocessing, such as normalizing and handling missing data, further ensure the model learns accurately and performs well on predictions [49]. Previous studies analyzing color images of the oral cavity primarily employed traditional machine learning models that relied on image preprocessing to extract specific features, such as color variation, shapes, and textures of oral lesions. In contrast to these traditional models, deep convolutional neural network (CNN) models can automatically extract distinctive features by examining the pixels of a color image, eliminating the need for manual feature extraction methods [50–52]. The robust learning capabilities of CNNs enable them to outperform traditional machine learning approaches, positioning them as an advanced technique in computer-aided cancer diagnosis [53]. Our study implemented a novel approach for convolutional neural networks (CNNs) using transfer models, specifically DenseNet-121, EfficientNet-B0, and ResNet-50, with the highest accuracy achieved by the DenseNet-121 algorithm. In studies by Warin et al. [29, 54], two deep learning models, DenseNet-121 and Faster R-CNN, were used to evaluate AUC, sensitivity, and specificity. They achieved sensitivity and specificity of 100%, with an AUC of 74% and 99%, surpassing our results. This difference can be attributed to their larger dataset, which included over 700 images, allowing for a greater number of cases than in our study and ultimately leading to improved results. On the other hand, Jubair et al. [30] employed the EfficientNet-B0 model for intraoral photos, and although their dataset (716 photos) was larger than ours, they reported a sensitivity of 87%, specificity of 85%, and an AUC of 93% in detecting early signs of oral cancer, comparable to our findings. These results highlight the consistency of AI algorithms in interpreting visual data for accurate diagnosis across different studies. Additionally, using ensemble voting techniques could further improve detection accuracy by combining multiple models with high precision [55]. Another method to enhance results is by modifying dynamic training, where feedback is provided to the machine on incorrect diagnoses during the process. This allows for error reduction and model refinement, improving performance for specific regions and populations [55]. Banavar et al. [31] highlighted the critical role of pathological imaging in leveraging artificial intelligence (AI) for diagnosing oral cancer. Their study focused on the analysis of histopathological slides using transfer learning techniques for early detection of the disease. The study reported remarkable outcomes, with ResNet-50 achieving a sensitivity and specificity of 97%. The focus on histopathological slides underscores the importance of microscopic examination in accurately diagnosing oral cancer lesions. These slides offer detailed insights into cellular and tissue changes, making them a valuable resource for training and validating AI algorithms. Several studies [12, 56–58] have also analyzed textual information, such as patient symptoms, medical history, and reported complaints, demonstrating that AI models are highly capable of detecting subtle patterns associated with the early symptoms of oral cancer. This approach not only complements existing imaging techniques but also offers a more comprehensive understanding of the patient’s clinical condition, contributing to a more robust and thorough diagnostic framework. One limitation of the study was the limited number of images, resulting from incomplete patient records and the lack of corresponding pathology reports, even when lesion images were available. So, international collaboration is needed to create large datasets of oral cancer and oral potentially malignant disorder (OPMD) clinical images, which would aid in the development and optimization of reliable CNNs for oral cancer detection. This study included a total of 518 clinical photographs, of which 104 (20%) represented malignant lesions and 414 (80%) benign or normal cases. This imbalance between classes may influence the generalizability of the trained models and should be acknowledged as a limitation. Moreover, clinical images diagnosed with epithelial dysplasia were excluded from the dataset. Although this decision was made to ensure diagnostic clarity, it limits the model’s applicability to cancer screening contexts where oral potentially malignant disorders are frequently encountered. Despite these limitations, many studies in related articles focus on specific types of oral lesions, such as mouth ulcers or tongue lesions, which represent only a small portion of oral conditions [59–61]. One of the strengths of our study is that we addressed this limitation by examining all types of lesions across the entire oral cavity. Our findings suggest that the proposed CNN model can be developed into a mobile application, offering a potentially practical, low-cost, and universally accessible solution for the early detection of oral cancer. Additionally, this approach could be utilized for remote monitoring of patients who have undergone surgery or treatment for disease recurrence [62]. Moreover, we recommend international collaboration with authors of similar studies to build a larger dataset of oral cancer and clinical images of oral potentially malignant disorders (OPMD). Conclusion The use of clinical images and artificial intelligence models offers numerous advantages in the diagnosis of oral cancer. These images provide a non-invasive tool that captures detailed information about lesion morphology, color variations, and structural changes, facilitating the early detection of abnormalities. However, it is important to recognize that while these AI-based approaches serve as valuable complementary methods, they should not be seen as a replacement for expert clinical diagnosis. Instead, they can enhance the accuracy and efficiency of early detection when used alongside professional evaluation. Moreover, AI should be viewed as a supportive technology that complements, rather than substitutes, expert clinical judgment. Acknowledgements Not applicable. Authors’ contributions Conceptualization: RR; Data collection: SSS, MM, FR, AS, ASh, NN; Data verification: SSS, MMG, FR, AS, ASh, NN; Formal analysis: SSS, MMG, FR, AS, ASh, NN; Methodology: SSS, MMG, FR, AS, ASh, NN; Visualization: SSS, MMGAS, ASh; Writing – original draft: SSS, FR, AS; Writing – review & editing: MMG, ASh, AS; Supervision: RR; Resources: RR; Project administration: RR. All authors have read and agreed to the final version for publication. Funding Not applicable. Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate The ethics committee of Kerman University of Medical Sciences approved this study (Ethical codes: ID: IR.KMU.REC.1402.238). Written informed consent was obtained from the participants. The participants’ data were kept confidential and were only accessible to the researchers. This study was conducted in accordance with the principles of the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement No datasets were generated or analysed during the current study.