A comparative study of single-stage and dual-stage classification models for OPMDs Authors: Jiayin Yu, Rui Huang, Xuan Wang, Yiran Zhang, Yueling Li, Yuhang Ye, Mei Ouyang, Weilan Su, Jing Wen, Zihe Zhang, Qinwen Liu, Zhiwen Liu, Wei Liu Published: 2025 ABSTRACT BackgroundEarly diagnosis of oral potentially malignant disorders (OPMDs) and oral cancer is crucial for reducing oral cancer incidence and mortality. With advancements in deep learning for oral image recognition, this study addressed the limitations of public datasets for oral mucosal disease, which are often restricted in size and insufficient in disease-type coverage. Through comparative analysis, a dual-stage multi-classification approach was proposed.MethodsThis study established and publicly released a high-quality oral mucosal disease dataset comprising 1,348 images, covering five categories: normal oral mucosa, oral leukoplakia, oral lichen planus, oral submucous fibrosis, and oral cancer. After image preprocessing, we trained and evaluated the models. Ten pre-trained models commonly used in oral image recognition (DenseNet-169, EfficientNet-B0, HRNet-W18-C, Inception-V4, MixNet-S, MobileNetV3-Large, ResNet-101, Swin Transformer, ViT-B, and YOLOv11l) were trained using two classification pathways: single-stage and dual-stage. Model performance was evaluated using metrics including accuracy, precision, recall, F1-score, area under the curve (AUC), and confusion matrix.ResultsThe dual-stage classification model based on Swin Transformer and DenseNet-169 achieved an accuracy of 0.9029, precision of 0.9082, recall of 0.8995, F1-score of 0.9032, and an AUC of 0.9735. Among the single-stage classification models, the best-performing EfficientNet-B0 model achieved an accuracy of 0.8710, precision of 0.8715, recall of 0.8719, F1-score of 0.8698, and an AUC of 0.9766. Based on these metrics, the dual-stage classification model demonstrated superior performance compared to the single-stage classification model.ConclusionsThis study established a publicly available high-quality dataset of 1,348 oral mucosal disease images. Furthermore, it proposed a novel dual-stage classification model integrating Swin Transformer and DenseNet-169, which was demonstrated to outperform conventional single-stage classification model across key performance metrics such as accuracy and precision. Introduction Oral cancer (OCA) refers to all malignant neoplasms that arise in the lip, buccal mucosa, hard and soft palate, floor of the mouth, gingiva, anterior two-thirds of the tongue and other unspecified parts of the mouth, with oral squamous cell carcinoma (OSCC) being the most common type [1]. Globally, it is estimated that there are 370,000 new cases of OCA and 170,000 OCA-related deaths each year [2], with a five-year mortality rate of approximately 50% [3]. This makes OCA a public health issue worthy of attention. Notably, OSCC often develops from oral potentially malignant disorders (OPMDs), primarily oral leukoplakia (OLK), oral lichen planus (OLP) and oral submucous fibrosis (OSF) [4]. Previous studies indicate that early detection and intervention of OPMDs can prevent malignant transformation in approximately 88% of OPMDs cases [5, 6], and early diagnosis of cancerous lesions can reduce cancer staging and significantly improve patient outcomes [7]. Thus, improving early detection and diagnosis is essential for reducing the incidence and mortality of OCA. Conventional diagnostic approaches for OCA include clinical oral examination (COE) and histopathological biopsy. However, the accuracy of COE is heavily dependent on the clinician expertise, introducing subjectivity, particularly in resource-limited settings where misdiagnosis is common. Although histopathology remains the gold standard, its invasive nature and potential for sampling bias limit its applicability [8]. These limitations underscore the need for an efficient, non-invasive adjunctive tool. Given the widespread adoption of smartphones, advancements in imaging technology, and the superficial location with direct visual accessibility of oral mucosa, the acquisition of oral images demonstrates significant feasibility and universal applicability. As both OCA and OPMDs exhibit morphological features that can be visually assessed, previous studies have demonstrated moderate-to-high concordance between specialist COE and the remote interpretation of oral photographs [9]. Therefore, an artificial intelligence (AI) system capable of matching the visual diagnostic proficiency of specialists holds excellent clinical application prospects. The rapid evolution of AI has driven notable progress in analyzing natural-light oral images using deep learning (DL) techniques [10]. Currently, binary classification tasks have yielded promising results. For instance, Warin et al. [11] employed a DenseNet121 model to classify a dataset of 350 normal mucosa images and 350 OSCC images, achieving an accuracy of 99% and a sensitivity of 98.75%. Fu et al. [12] proposed a cascaded convolutional neural network (CNN) that localized lesions via object detection before performing classification. This dual-stage framework achieved a sensitivity of 97.4% for early OSCC detection, marking a significant advance in clinical utility. Using an Inception-v3 architecture, Keser et al. [13] attained 100% classification accuracy for OLP images, further validating AI’s efficacy in diagnosing specific oral diseases. Substantial progress has also been made in the multi-class classification of oral mucosal diseases. In experiments on a three-class dataset (Normal, OPMDs, and OCA), models including ResNet-152, DenseNet-161, Inception-V4, and EfficientNet-B4 were compared. Results indicated that EfficientNet-B4 and Inception-V4 were the most promising, with an F1-score of 85.8% [14]. As for object detection and image classification of OLK, OLP and OSCC, both Vision Transformer model and YOLOX-based CNN model achieved precision above 85% and specificity up to 95%. The latter model reached a level comparable to the specialists of oral mucosal disease in multi-center clinical evaluations [15, 16]. Although existing models have achieved preliminary success in multi-classification tasks for oral mucosal diseases, there still remain limitations such as insufficiently detailed disease coverage, suboptimal accuracy, and inability to fully meet the needs of precise clinical diagnosis and treatment. This study contributes to the field in the following three key areas: We established and publicly released a high-quality oral mucosal disease dataset with relatively balanced image counts across five categories.The study accomplished a five-classification task, encompassing normal oral mucosa, three common OPMDs—OLK, OLP, and OSF—as well as OCA.We proposed a novel dual-stage classification model integrating Swin Transformer and DenseNet-169, which achieved a 3.19% higher accuracy than the best-performing single-stage classification model. We established and publicly released a high-quality oral mucosal disease dataset with relatively balanced image counts across five categories. We established and publicly released a high-quality oral mucosal disease dataset with relatively balanced image counts across five categories. The study accomplished a five-classification task, encompassing normal oral mucosa, three common OPMDs—OLK, OLP, and OSF—as well as OCA. The study accomplished a five-classification task, encompassing normal oral mucosa, three common OPMDs—OLK, OLP, and OSF—as well as OCA. We proposed a novel dual-stage classification model integrating Swin Transformer and DenseNet-169, which achieved a 3.19% higher accuracy than the best-performing single-stage classification model. We proposed a novel dual-stage classification model integrating Swin Transformer and DenseNet-169, which achieved a 3.19% higher accuracy than the best-performing single-stage classification model. Materials and methodology Establishment of dataset This study compiled a dataset of 1,348 intraoral mucosal images categorized into five classes: Normal oral mucosa, OLK, OLP, OSF, and OCA. As shown in Table 1; Fig. 1, the image count for each category ranges from 225 to 310 (16.69%–23.00%), indicating a relatively balanced distribution. Annotations were performed by four oral mucosal disease specialists, each possessing over ten years of clinical experience. To ensure label reliability, cases with initial diagnostic disagreements were first discussed in a consensus meeting. Images for which a consensus could not be reached were excluded from the final dataset. However, it should be acknowledged that the final labels were based solely on clinical evaluation in the absence of histopathological confirmation. Images of normal mucosa were captured from various anatomical sites and angles, while the other four types of oral images clearly displayed surface features of the lesions (Fig. 2). These lesions were located on diverse oral sites, including the lip, buccal mucosa, hard and soft palate, floor of the mouth, gingiva, and anterior two-thirds of the tongue. The dataset exhibited significant variability in quality (including lighting, scaling, angle, and clarity) and resolution. This heterogeneity enhances the dataset’s applicability to real-world scenarios and provides valuable samples for training classification models with robust generalization capabilities. Table 1Image counts for the five diagnostic categories in the oral mucosal disease datasetImage categoriesCountsNormal Oral Mucosa286OLK253OLP310OSF225OCA274 The majority of images were collected from three hospitals in China: Changsha Stomatological Hospital, the Second Xiangya Hospital of Central South University, and the First Affiliated Hospital of Hunan University of Traditional Chinese Medicine. These images were captured using various types of smartphones and digital cameras. A small number of images were sourced from https://openi.nlm.nih.gov/gridquery?m=1&n=100&it=xg&q=oral, endowing the dataset with a polycentric characteristic. Table 1Image counts for the five diagnostic categories in the oral mucosal disease datasetImage categoriesCountsNormal Oral Mucosa286OLK253OLP310OSF225OCA274 Image counts for the five diagnostic categories in the oral mucosal disease dataset Fig. 1Distribution of images in the oral mucosal disease dataset across five categories Distribution of images in the oral mucosal disease dataset across five categories Fig. 2Sample images from the five categories: (a) Normal Oral Mucosa (b) OLK (c) OLP (d) OSF (e) OCA Sample images from the five categories: (a) Normal Oral Mucosa (b) OLK (c) OLP (d) OSF (e) OCA All images in the dataset were stored in JPG format and organized into separate directories labeled Normal, OLK, OLP, OSF, and OCA. The naming convention for image files was ‘category number-image number’. The directory names corresponded to the assigned category numbers (1–5), specifically Normal (1), OLK (2), OLP (3), OSF (4), and OCA (5). This naming scheme struck a balance between simplicity and organization, providing clear labels and an efficient framework for data management during model training. This study was approved by the Ethics Committee of Changsha Stomatological Hospital (Ethics Number: 2025 − 169). All data were anonymized to exclude patients’ personal and private information. Data preprocessing All images were resized to 224 × 224 pixels and randomly partitioned into training, validation, and test sets at an 8:1:1 ratio. To mitigate the risk of overfitting resulting from the limited dataset size, data augmentation was applied exclusively to the training set. The augmentation techniques included rotation (90°, 180°, 270°), horizontal flipping, horizontal and vertical translation (by 10% of the image size), and color space adjustments (brightness = 0.5, contrast = 0.1, saturation = 0.01) [17]. Model selection This study employed ten pre-trained models encompassing a spectrum of classical architectures, which are widely adopted and have demonstrated strong performance in oral mucosal disease image recognition. The selected models are described below. DenseNet-169 is a 169-layer CNN characterized by its dense connectivity, where each layer receives input from all preceding layers. This promotes feature reuse, mitigates the vanishing gradient problem, and reduces model parameters. The architecture comprises dense blocks connected by transition layers for dimensionality reduction, ending with a global average pooling and a linear layer for classification [18]. EfficientNet-B0 utilizes a compound scaling method to systematically balance the scaling of depth, width, and resolution via a single coefficient [19]. The core architecture consists of MBConv blocks, further integrated with Squeeze-and-Excitation (SE) modules and the Swish activation function to enhance model capacity and feature learning [20, 21]. Across its 18-layer architecture, HRNet-W18-C maintains multi-resolution representations by simultaneously processing feature maps at multiple scales through parallel branches. The architecture employs cross-resolution bidirectional fusion modules and a reduction in the channel width of high-resolution representations to form a lightweight design [22]. Inception-V4 enhances the Inception family through refined architecture, employing a deep stem for initial feature extraction, factorized Inception modules for efficient multi-scale processing, and specialized reduction blocks for controlled downsampling. This design achieves an improved balance between accuracy and computational cost [23]. MixNet-S employs mixed depthwise convolution. This operation partitions channels into groups that are processed by kernels of different sizes, enabling the efficient capture of multi-scale features. This design allows MixNet-S to achieve a competitive 75.8% top-1 accuracy on ImageNet with a computational cost of only 256 M FLOPs, striking an excellent balance between efficiency and performance [24]. MobileNetV3-Large was co-designed by neural architecture search and manual refinement, integrating a lightweight SE module and the hardware-efficient h-swish activation function. This approach achieves a strong balance between computational cost and model accuracy [25]. ResNet-101 addresses deep network training challenges through two core designs: a bottleneck structure for computational efficiency, and a residual framework with skip connections to mitigate vanishing gradients. Stacking these blocks enables stable training and hierarchical feature construction, preventing performance degradation despite increased depth [26]. Swin Transformer is a state-of-the-art vision model that introduces a shifted window self-attention mechanism as its core innovation. This mechanism limits computational complexity by confining attention to non-overlapping local windows. By shifting these windows between layers, it efficiently models global context, achieving an optimal balance between accuracy and efficiency [27]. ViT-B is a pure Transformer architecture for computer vision. It operates by splitting images into patches, adding positional embeddings, and processing them via global self-attention. While this enables powerful visual relationship modeling, it makes the model data-intensive. Consequently, ViT-B excels when pre-trained on large datasets but underperforms ResNet on small-scale data [28]. YOLOv11 introduces an optimized CSPDarknet backbone that replaces C2F with efficient C3K2 modules, reducing parameters and computation while preserving the receptive field. It further incorporates a C2PSA module after SPPF to enhance multi-scale feature extraction, significantly boosting small object detection performance [29]. Model parameter settings The experimental training was conducted on a workstation equipped with an NVIDIA GeForce RTX 3090 GPU. For model training, we initially set a maximum of 100 epochs. An early-stopping strategy was implemented to prevent overfitting: the training process would terminate prematurely if the validation accuracy did not improve for 10 consecutive epochs. The best-performing model was selected based on the highest validation accuracy achieved during training. Table 2 shows the optimal model parameters obtained through experiments, including batch size, learning rate, and optimizer for different models. Table 2Model parametersModelBatch sizeLearning rateOptimizerDenseNet-169160.001SGDEfficientNet-B0160.01SGDHRNet-W18-C320.05SGDInception-V41280.1SGDMixNet-S80.001SGDMobileNetV3-Large320.001SGDResNet101160.0001AdamSwin Transformer80.0001AdamWViT-B80.001SGDYOLOv11l320.001AdamW Table 2Model parametersModelBatch sizeLearning rateOptimizerDenseNet-169160.001SGDEfficientNet-B0160.01SGDHRNet-W18-C320.05SGDInception-V41280.1SGDMixNet-S80.001SGDMobileNetV3-Large320.001SGDResNet101160.0001AdamSwin Transformer80.0001AdamWViT-B80.001SGDYOLOv11l320.001AdamW Model parameters Model architecture Figure 3 illustrates the overall workflow of this study. Following data collection and preprocessing, images are processed through two parallel approaches for model training and evaluation. Fig. 3Overall workflow Overall workflow The first approach employed a single-stage five-classification strategy. It utilized ten pre-trained DL models to classify oral images into one of five categories: Normal, OLK, OLP, OSF or OCA. The second approach employed a pipeline-based dual-stage five-classification strategy. In the first stage, the same ten pre-trained models were trained and evaluated to differentiate among three categories: Normal, OPMDs, and OCA. In the second stage, the models were trained to subclassify OPMDs into OLK, OLP, and OSF. The best-performing models from each stage were then integrated into a conditional workflow: an input image first undergoes initial classification by the first-stage model; if predicted as OPMDs, it is forwarded to the second-stage model for fine-grained classification into one of the three OPMD subtypes. Otherwise, the initial prediction (Normal or OCA) is taken as the final output. The overall performance of both strategies was evaluated and compared. Assessment metrics To assess the performance of all models, we used the following metrics on true positive (TP), true negative (TN), false positive (FP) and false negative (FN): Accuracy is the most intuitive performance assessment metric, indicating the number of samples correctly predicted by the model as a proportion of the total number of samples.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}Accuracy=\frac{TP+TN}{TP+TN+FP+FN}\#\end{array}$$\end{document} Precision is used to evaluate the ratio of correctly predicted observations of each category. It is concerned with the accuracy of the predicted results.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}Precision=\frac{TP}{TP+FP}\#\end{array}$$\end{document} Recall indicates the proportion of all samples that are positive categories that are correctly predicted as positive by the model. It is concerned with the model’s ability to capture positive categories.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}Recall=\frac{TP}{TP+FN}\#\end{array}$$\end{document} F1-score is the reconciled average of precision and recall and is used to measure the comprehensive performance of the model. When there is a large gap between precision and recall, the F1-score can provide a balanced evaluation.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}F1-score=2\times\:\frac{Precision\times\:Recall}{Precision+Recall}\#\end{array}$$\end{document} Confusion Matrix is a table used to represent the relationship between the predicted and actual results of a visualization model. It usually contains four components: TP, FN, FP and TN. Area Under the Curve (AUC): The Receiver Operating Characteristic (ROC) curve is a graphical tool for evaluating the performance of a model by plotting the true positive rate against the false positive rate at different thresholds. AUC denotes the area under the ROC curve and has a value between 0 and 1, with larger values indicating better model performance. Results and discussion In this study, a comparative analysis was first conducted on 1,348 oral images using ten DL classification models (Table 3). To ensure the robustness of the results, the entire experiment was repeated five times using the same training/validation/test set split. Cross-validation was not employed in this study. The final performance metrics are reported as the mean ± standard deviation across these five independent runs. Table 3Performance comparison of single-stage classification modelsModelAccuracyPrecisionRecallF1-scoreAUCEfficientNet-B00.8710 ± 0.01570.8715 ± 0.01680.8719 ± 0.01480.8698 ± 0.01600.9766 ± 0.0046DenseNet-1690.8652 ± 0.00400.8701 ± 0.00680.8654 ± 0.00420.8650 ± 0.00430.9756 ± 0.0049Swin Transformer0.8652 ± 0.03030.8673 ± 0.03150.8654 ± 0.03120.8648 ± 0.03240.9761 ± 0.0062Inception-V40.8478 ± 0.02680.8564 ± 0.02560.8478 ± 0.02680.8478 ± 0.02690.9781 ± 0.0109YOLOv11l0.8376 ± 0.00410.8428 ± 0.00320.8409 ± 0.00510.8400 ± 0.00500.9643 ± 0.0010ResNet1010.8087 ± 0.04370.8265 ± 0.02290.8101 ± 0.04210.8006 ± 0.05230.9633 ± 0.0117ViT-B0.6493 ± 0.02920.6521 ± 0.02440.6427 ± 0.02940.6350 ± 0.03280.8912 ± 0.0147HRNet-W18-C0.5841 ± 0.08130.5913 ± 0.06000.5857 ± 0.07640.5767 ± 0.08340.8517 ± 0.0379MixNet-S0.5595 ± 0.04220.5923 ± 0.02380.5668 ± 0.04110.5580 ± 0.04360.8409 ± 0.0160MobileNetV3-Large0.4841 ± 0.07600.5031 ± 0.04780.4832 ± 0.07240.4754 ± 0.07320.7713 ± 0.0509 Table 3Performance comparison of single-stage classification modelsModelAccuracyPrecisionRecallF1-scoreAUCEfficientNet-B00.8710 ± 0.01570.8715 ± 0.01680.8719 ± 0.01480.8698 ± 0.01600.9766 ± 0.0046DenseNet-1690.8652 ± 0.00400.8701 ± 0.00680.8654 ± 0.00420.8650 ± 0.00430.9756 ± 0.0049Swin Transformer0.8652 ± 0.03030.8673 ± 0.03150.8654 ± 0.03120.8648 ± 0.03240.9761 ± 0.0062Inception-V40.8478 ± 0.02680.8564 ± 0.02560.8478 ± 0.02680.8478 ± 0.02690.9781 ± 0.0109YOLOv11l0.8376 ± 0.00410.8428 ± 0.00320.8409 ± 0.00510.8400 ± 0.00500.9643 ± 0.0010ResNet1010.8087 ± 0.04370.8265 ± 0.02290.8101 ± 0.04210.8006 ± 0.05230.9633 ± 0.0117ViT-B0.6493 ± 0.02920.6521 ± 0.02440.6427 ± 0.02940.6350 ± 0.03280.8912 ± 0.0147HRNet-W18-C0.5841 ± 0.08130.5913 ± 0.06000.5857 ± 0.07640.5767 ± 0.08340.8517 ± 0.0379MixNet-S0.5595 ± 0.04220.5923 ± 0.02380.5668 ± 0.04110.5580 ± 0.04360.8409 ± 0.0160MobileNetV3-Large0.4841 ± 0.07600.5031 ± 0.04780.4832 ± 0.07240.4754 ± 0.07320.7713 ± 0.0509 Performance comparison of single-stage classification models In the single-stage classification model, EfficientNet-B0 demonstrated the highest overall performance. It achieved mean values of 0.8710 (accuracy), 0.8715 (precision), 0.8719 (recall), 0.8698 (F1-score), and 0.9766 (AUC) over five runs. While its AUC is marginally lower than that of Inception-V4, all other metrics ranked highest among the ten evaluated models (Table 3). This superior performance is attributed to EfficientNet-B0’s compound-scaling strategy, which delivers optimal computational efficiency by markedly reducing the parameter count while maintaining outstanding accuracy within the same architectural family [30]. Notably, the model demonstrates limited discriminative capability for OLK and OCA (Fig. 6), with class-specific accuracies of only 0.78 and 0.80, respectively. Consequently, based on the observed variability in feature learning across models, a dual-stage classification strategy was implemented. This framework simplifies the learning objective at each stage, which stands in contrast to prevalent dual-stage classification strategy that typically cascades an object detection or segmentation model with a classifier. Instead of first localizing lesion regions, our approach employs a hierarchical classification. The first-stage model is dedicated to differentiating the three major categories—Normal, OPMDs, and OCA—allowing it to concentrate on their core discriminative features. Subsequently, for cases identified as OPMDs, a specialized second-stage model is employed to discern the more subtle feature differences between OLK, OLP, and OSF subcategories. For the first stage of the dual-stage classification strategy, ten DL models were trained and evaluated on the triple-classification task (Normal, OPMDs, and OCA). Swin Transformer emerged as the optimal model, achieving superior five-run mean values for accuracy (0.9159), precision (0.9097), recall (0.9038), F1-score (0.9056), and AUC (0.9697), as detailed in Table 4. Its strong performance is attributed to the shifted-window attention mechanism [27], which excels at capturing global contextual information by efficiently modeling long-range dependencies. Table 4Performance comparison of dual-stage classification models (Stage 1)ModelAccuracyPrecisionRecallF1-scoreAUCSwin Transformer0.9159 ± 0.01820.9097 ± 0.02220.9038 ± 0.02040.9056 ± 0.01970.9697 ± 0.0079ResNet1010.8942 ± 0.01320.8909 ± 0.01760.8749 ± 0.02930.8801 ± 0.01760.9614 ± 0.0044DenseNet-1690.8870 ± 0.02150.8786 ± 0.02930.8671 ± 0.02100.8703 ± 0.02390.9545 ± 0.0063Inception-V40.8739 ± 0.03720.8770 ± 0.03100.8739 ± 0.03720.8703 ± 0.04240.9524 ± 0.0185EfficientNet-B00.8768 ± 0.04560.8656 ± 0.05160.8634 ± 0.04730.8622 ± 0.05030.9549 ± 0.0204YOLOv11l0.8565 ± 0.00790.8548 ± 0.02330.8086 ± 0.00520.8210 ± 0.00010.9523 ± 0.0060ViT-B0.7696 ± 0.03460.7611 ± 0.02740.7124 ± 0.03280.7274 ± 0.03020.8798 ± 0.0282MixNet-S0.7159 ± 0.03130.6893 ± 0.06130.6231 ± 0.05050.6231 ± 0.05050.8064 ± 0.0366HRNet-W18-C0.6812 ± 0.04380.6394 ± 0.04960.5699 ± 0.08870.5775 ± 0.08780.8131 ± 0.0501MobileNetV3-Large0.6348 ± 0.01960.5960 ± 0.05970.5414 ± 0.04180.5442 ± 0.04010.7237 ± 0.0234 Table 4Performance comparison of dual-stage classification models (Stage 1)ModelAccuracyPrecisionRecallF1-scoreAUCSwin Transformer0.9159 ± 0.01820.9097 ± 0.02220.9038 ± 0.02040.9056 ± 0.01970.9697 ± 0.0079ResNet1010.8942 ± 0.01320.8909 ± 0.01760.8749 ± 0.02930.8801 ± 0.01760.9614 ± 0.0044DenseNet-1690.8870 ± 0.02150.8786 ± 0.02930.8671 ± 0.02100.8703 ± 0.02390.9545 ± 0.0063Inception-V40.8739 ± 0.03720.8770 ± 0.03100.8739 ± 0.03720.8703 ± 0.04240.9524 ± 0.0185EfficientNet-B00.8768 ± 0.04560.8656 ± 0.05160.8634 ± 0.04730.8622 ± 0.05030.9549 ± 0.0204YOLOv11l0.8565 ± 0.00790.8548 ± 0.02330.8086 ± 0.00520.8210 ± 0.00010.9523 ± 0.0060ViT-B0.7696 ± 0.03460.7611 ± 0.02740.7124 ± 0.03280.7274 ± 0.03020.8798 ± 0.0282MixNet-S0.7159 ± 0.03130.6893 ± 0.06130.6231 ± 0.05050.6231 ± 0.05050.8064 ± 0.0366HRNet-W18-C0.6812 ± 0.04380.6394 ± 0.04960.5699 ± 0.08870.5775 ± 0.08780.8131 ± 0.0501MobileNetV3-Large0.6348 ± 0.01960.5960 ± 0.05970.5414 ± 0.04180.5442 ± 0.04010.7237 ± 0.0234 Performance comparison of dual-stage classification models (Stage 1) For the second stage of the dual-stage classification strategy, the ten DL models were trained and evaluated on the triple-classification task of distinguishing OLK, OLP, and OSF. DenseNet-169 emerged as the top performer; although recall and AUC lag behind EfficientNet-B0, other three metrics ranked highest among all models (Table 5). This advantage can be attributed to the cross-layer feature reuse mechanism [18], the densely connected architecture effectively captures local textural nuances and enables precise discrimination of OPMDs. Table 5Performance comparison of dual-stage classification models (Stage 2)ModelAccuracyPrecisionRecallF1-scoreAUCDenseNet-1690.9050 ± 0.01900.9162 ± 0.01250.9044 ± 0.01870.9070 ± 0.01780.9767 ± 0.0047EfficientNet-B00.9050 ± 0.02090.9071 ± 0.01950.9072 ± 0.02020.9057 ± 0.01980.9806 ± 0.0061ResNet1010.8850 ± 0.02710.8926 ± 0.02350.8854 ± 0.02730.8873 ± 0.02630.9659 ± 0.0150Swin Transformer0.8850 ± 0.03240.8957 ± 0.03360.8855 ± 0.03200.8859 ± 0.03400.9754 ± 0.0048YOLOv11l0.8725 ± 0.04540.8741 ± 0.04620.8787 ± 0.04380.8758 ± 0.04510.9751 ± 0.0065Inception-V40.8025 ± 0.16660.8412 ± 0.11050.8009 ± 0.17080.7871 ± 0.19970.8981 ± 0.1112ViT-B0.6925 ± 0.00680.7665 ± 0.00780.6983 ± 0.00800.6986 ± 0.00710.8515 ± 0.0028MixNet-S0.6450 ± 0.04970.6552 ± 0.05540.6528 ± 0.04830.6528 ± 0.04830.8083 ± 0.0380MobileNetV3-Large0.5825 ± 0.05900.6054 ± 0.05540.5809 ± 0.06190.5798 ± 0.05630.7498 ± 0.0462HRNet-W18-C0.5625 ± 0.05590.5748 ± 0.05740.5681 ± 0.06090.5606 ± 0.05400.7612 ± 0.0433 Table 5Performance comparison of dual-stage classification models (Stage 2)ModelAccuracyPrecisionRecallF1-scoreAUCDenseNet-1690.9050 ± 0.01900.9162 ± 0.01250.9044 ± 0.01870.9070 ± 0.01780.9767 ± 0.0047EfficientNet-B00.9050 ± 0.02090.9071 ± 0.01950.9072 ± 0.02020.9057 ± 0.01980.9806 ± 0.0061ResNet1010.8850 ± 0.02710.8926 ± 0.02350.8854 ± 0.02730.8873 ± 0.02630.9659 ± 0.0150Swin Transformer0.8850 ± 0.03240.8957 ± 0.03360.8855 ± 0.03200.8859 ± 0.03400.9754 ± 0.0048YOLOv11l0.8725 ± 0.04540.8741 ± 0.04620.8787 ± 0.04380.8758 ± 0.04510.9751 ± 0.0065Inception-V40.8025 ± 0.16660.8412 ± 0.11050.8009 ± 0.17080.7871 ± 0.19970.8981 ± 0.1112ViT-B0.6925 ± 0.00680.7665 ± 0.00780.6983 ± 0.00800.6986 ± 0.00710.8515 ± 0.0028MixNet-S0.6450 ± 0.04970.6552 ± 0.05540.6528 ± 0.04830.6528 ± 0.04830.8083 ± 0.0380MobileNetV3-Large0.5825 ± 0.05900.6054 ± 0.05540.5809 ± 0.06190.5798 ± 0.05630.7498 ± 0.0462HRNet-W18-C0.5625 ± 0.05590.5748 ± 0.05740.5681 ± 0.06090.5606 ± 0.05400.7612 ± 0.0433 Performance comparison of dual-stage classification models (Stage 2) Consequently, an integrated dual-stage classification model was constructed by linking the optimal first-stage model (Swin Transformer) with the optimal second-stage model (DenseNet-169) via a conditional execution pathway. The overall dual-stage classification model achieves an accuracy of 0.9029, precision of 0.9082, recall of 0.8995, F1-score of 0.9032, and an AUC of 0.9735. Compared with the three top-performing single-stage classification models, the dual-stage classification model shows higher accuracy, precision, recall and F1-score (Fig. 4); A paired t-test comparing the direct results of five runs of the best single-stage classification model, which is EfficientNet-B0, with those of the dual-stage classification model yielded t-test p-values of 0.040, 0.044, 0.048, and 0.045 for accuracy, precision, recall, and F1-score, respectively. All these values are below 0.05, indicating a statistically significant improvement. However, the paired t-test p-value for the AUC comparison is 0.417, indicating no significant difference compared to EfficientNet-B0. Fig. 4Performance comparison of single-stage and dual-stage classification models Performance comparison of single-stage and dual-stage classification models The ROC curves (Fig. 5) demonstrate that both the single-stage EfficientNet-B0 and the dual-stage classification model performed well across all categories (AUC > 0.95), with the curves closely approaching the upper left-hand corner, confirming high discriminatory power, suggesting strong potential for clinical application. Fig. 5ROC curves (a) Single-stage classification model based on EfficientNet-B0 (b) Dual-stage classification model based on Swin Transformer and DenseNet-169 ROC curves (a) Single-stage classification model based on EfficientNet-B0 (b) Dual-stage classification model based on Swin Transformer and DenseNet-169 Figure 6 presents the average confusion matrices from five experimental runs, comparing the performance of single-stage EfficientNet-B0 model and the dual-stage classification model. The dual-stage classification model achieved higher recognition accuracy than the single-stage classification model across all four categories (Normal, OCA, OLK, and OLP). Notably, the dual-stage classification model achieved 99% accuracy in identifying normal oral mucosa, demonstrating strong capability to distinguish between normal and pathological oral mucosa. Furthermore, the dual-stage classification model demonstrated significant improvement in OLP recognition accuracy, increasing from 0.87 to 0.94. However, its recognition accuracy for OCA and OLK remains relatively low at 0.82 and 0.80 respectively, with an improvement of only 2%, implying the risk of missed malignancy in clinical practice. This limited enhancement may be attributed to the textural similarities between early-stage OCA and OPMDs, suggesting an area for future model refinement. Fig. 6Confusion matrix (a) Single-stage classification model based on EfficientNet-B0 (b) Dual-stage classification model based on Swin Transformer and DenseNet-169 Confusion matrix (a) Single-stage classification model based on EfficientNet-B0 (b) Dual-stage classification model based on Swin Transformer and DenseNet-169 In summary, the comprehensive performance of the dual-stage classification model is better than that of the single-stage classification model, but it suffers from the disadvantage of relatively low OCA and OLK recognition accuracy. To interpret the model’s decisions, we generated class-activation maps using Gradient-weighted Class Activation Mapping (Grad-CAM) [31]. The colour heatmap highlights the areas in the input image contributing to the decision made by the model, with red and yellow regions representing a high score for the class. Grad-CAM shows that the dual-stage classification model effectively focuses on predicting the relevant areas (Fig. 7). To quantitatively assess the segmentation performance, we used the Intersection over Union (IoU) metric to measure the overlap between the predicted segmentation mask and the ground-truth mask. The IoU is calculated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:IoU=\frac{\left|P\cap\:G\right|}{\left|P\cup\:G\right|}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P$$\end{document} is the predicted lesion region and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G$$\end{document} is the ground-truth lesion region. A higher IoU value (closer to 1) indicates better segmentation performance. For Fig. 7(a), the predicted segmentation mask (d) exhibits a high overlap with the ground truth annotation (b), achieving an IoU value of 0.7372. In contrast, for Fig. 7(e), the predicted segmentation mask (h) demonstrates a lower overlap with the ground truth annotation (f), yielding an IoU value of 0.5811. This indicates that the image segmentation capability of the dual-stage classification model is unstable, and its generalization ability requires improvement. Future work will focus on enhancing spatial consistency by incorporating region of interest (ROI) labels during training to better align the activation maps with precise lesion boundaries. Fig. 7Grad-CAM visualization. (a)(e) Original image (b)(f) Ground-truth lesion segmentation masks (c)(g) Grad-CAM heatmaps (d)(h) Lesion segmentation image derived from Grad-CAM heatmaps Grad-CAM visualization. (a)(e) Original image (b)(f) Ground-truth lesion segmentation masks (c)(g) Grad-CAM heatmaps (d)(h) Lesion segmentation image derived from Grad-CAM heatmaps To comprehensively evaluate the performance of various pre-trained DL classification models in detecting oral mucosal lesions, we systematically reviewed and compared relevant studies from the current literature, as summarized in Table 6. Table 6Review and comparison of pre-trained DL modelsAuthorPre-trained DL classification models used for experimentationDatasetClassification tasksThe best performing modelResults Lin et al. 2021 [32]VGG16, ResNet50, DenseNet169, HRNet-W18688 images of oral lesions and 760 images of normal mucosa collected retrospectively using four different smartphonesNormal, aphthous ulcer, low-risk OPMDs, high-risk OPMDs, and OCAHRNet-W18Sensitivity:83.0%Specificity:96.6%Precision:84.3%F1-score:83.6% Tanriver et al. 2021 [14]ResNet-152, DenseNet-161, Inception-V4, EfficientNet-B4684 images of oral lesions collected from the Oncology Institute and the InternetBenign, OPMDs, and OCAInception-v4Accuracy:87.7%Recall: 85.5%F1-score: 85.8% Sharma et al. 2022 [6]MobileNet, VGG19, VGG16, Inception-V3, ResNet50106 normal images, 102 OPMDs images, and 121 OSCC images acquired using smartphonesNormal, OPMDs, and OSCCVGG19Accuracy:76% Warin et al. 2022 [33]DenseNet-169, ResNet-101, SqueezeNet, Swin-S365 OSCC, 315 OPMDs, and 300 non-pathological clinical oral photographic imagesNormal, OPMDs, and OSCCDenseNet-196OSCC:Precision:0.98Sensitivity:0.99OPMDs:Precision:0.95Sensitivity:0.95 Duhayyim et al. 2022 [34]ADCOL-DL, Inception-V4, DenseNet, RF, ANN-SVM, C-Net, CADOC-SFOFC87 cancer images and 44 non-cancer images from Kaggle repositoryCancer and non-cancerCADOC-SFOFCAccuracy:98.11%Precision:98.72%Recall:96.67%F1-score: 97.63% Mira et al. 2024 [35]HRNet-W18,VGG-16,ResNet-50,DenseNet-169760 normal mucosa, 251 aphthous ulcer, 231 low-risk OPMDs, 141 high-risk OPMDs, and 65 OCA acquired using the mobile phone’s built-in camera application.Normal, aphthous ulcer, low-risk OPMDs, high-risk OPMDs, and OCAHRNet-W18Accuracy:84.3%Sensitivity:83.0%Specificity:96.0%F1-score: 83.6% Rabinovici-Cohen et al. 2024 [36]CNN (2D ResNet-18),OCNN(ResNet50),CNN Meta,OCNN Meta,CNN Meta + ISIC Pretrained1382 photos derived from the selected 586 patients at TZMC, 1016 photos obtained from 884 patients selected from the InternetNormal, Benign, Mimic-cancer, and OSCCCNN MetaAUC:86.0%Sensitivity:96.0%Our WorkEfficientNet-B0, YOLOv11l, ViT-B, Swin Transformer, Inception-V4, DenseNet-169, MobileNetV3-Large, HRNet-W18-C, MixNet-S,ResNet-101, Dual-Stage Classification Model (Swin-Transformer + DenseNet-169)1348 oral images collected by smartphones, digital cameras and literatureNormal, OLP, OLK, OSF, and OSCCDual-Stage Classification ModelAccuracy:90.29%Precision:90.82%Recall:89.95%F1-score:90.32%AUC:97.35% Table 6Review and comparison of pre-trained DL modelsAuthorPre-trained DL classification models used for experimentationDatasetClassification tasksThe best performing modelResults Lin et al. 2021 [32]VGG16, ResNet50, DenseNet169, HRNet-W18688 images of oral lesions and 760 images of normal mucosa collected retrospectively using four different smartphonesNormal, aphthous ulcer, low-risk OPMDs, high-risk OPMDs, and OCAHRNet-W18Sensitivity:83.0%Specificity:96.6%Precision:84.3%F1-score:83.6% Tanriver et al. 2021 [14]ResNet-152, DenseNet-161, Inception-V4, EfficientNet-B4684 images of oral lesions collected from the Oncology Institute and the InternetBenign, OPMDs, and OCAInception-v4Accuracy:87.7%Recall: 85.5%F1-score: 85.8% Sharma et al. 2022 [6]MobileNet, VGG19, VGG16, Inception-V3, ResNet50106 normal images, 102 OPMDs images, and 121 OSCC images acquired using smartphonesNormal, OPMDs, and OSCCVGG19Accuracy:76% Warin et al. 2022 [33]DenseNet-169, ResNet-101, SqueezeNet, Swin-S365 OSCC, 315 OPMDs, and 300 non-pathological clinical oral photographic imagesNormal, OPMDs, and OSCCDenseNet-196OSCC:Precision:0.98Sensitivity:0.99OPMDs:Precision:0.95Sensitivity:0.95 Duhayyim et al. 2022 [34]ADCOL-DL, Inception-V4, DenseNet, RF, ANN-SVM, C-Net, CADOC-SFOFC87 cancer images and 44 non-cancer images from Kaggle repositoryCancer and non-cancerCADOC-SFOFCAccuracy:98.11%Precision:98.72%Recall:96.67%F1-score: 97.63% Mira et al. 2024 [35]HRNet-W18,VGG-16,ResNet-50,DenseNet-169760 normal mucosa, 251 aphthous ulcer, 231 low-risk OPMDs, 141 high-risk OPMDs, and 65 OCA acquired using the mobile phone’s built-in camera application.Normal, aphthous ulcer, low-risk OPMDs, high-risk OPMDs, and OCAHRNet-W18Accuracy:84.3%Sensitivity:83.0%Specificity:96.0%F1-score: 83.6% Rabinovici-Cohen et al. 2024 [36]CNN (2D ResNet-18),OCNN(ResNet50),CNN Meta,OCNN Meta,CNN Meta + ISIC Pretrained1382 photos derived from the selected 586 patients at TZMC, 1016 photos obtained from 884 patients selected from the InternetNormal, Benign, Mimic-cancer, and OSCCCNN MetaAUC:86.0%Sensitivity:96.0%Our WorkEfficientNet-B0, YOLOv11l, ViT-B, Swin Transformer, Inception-V4, DenseNet-169, MobileNetV3-Large, HRNet-W18-C, MixNet-S,ResNet-101, Dual-Stage Classification Model (Swin-Transformer + DenseNet-169)1348 oral images collected by smartphones, digital cameras and literatureNormal, OLP, OLK, OSF, and OSCCDual-Stage Classification ModelAccuracy:90.29%Precision:90.82%Recall:89.95%F1-score:90.32%AUC:97.35% Review and comparison of pre-trained DL models Sensitivity:83.0% Specificity:96.6% Precision:84.3% F1-score:83.6% Accuracy:87.7% Recall: 85.5% F1-score: 85.8% OSCC: Precision:0.98 Sensitivity:0.99 OPMDs: Precision:0.95 Sensitivity:0.95 Accuracy:98.11% Precision:98.72% Recall:96.67% F1-score: 97.63% HRNet-W18, VGG-16, ResNet-50, DenseNet-169 Accuracy:84.3% Sensitivity:83.0% Specificity:96.0% F1-score: 83.6% CNN (2D ResNet-18), OCNN(ResNet50), CNN Meta, OCNN Meta, CNN Meta + ISIC Pretrained AUC:86.0% Sensitivity:96.0% EfficientNet-B0, YOLOv11l, ViT-B, Swin Transformer, Inception-V4, DenseNet-169, MobileNetV3-Large, HRNet-W18-C, MixNet-S, ResNet-101, Dual-Stage Classification Model (Swin-Transformer + DenseNet-169) Accuracy:90.29% Precision:90.82% Recall:89.95% F1-score:90.32% AUC:97.35% As shown in Table 6, different studies have demonstrated that when using pre-trained DL models for oral lesion recognition, the model performance is affected by multiple factors such as dataset size, image quality, classification tasks, and model architecture. Sharma et al. [6] achieved only 76% accuracy with their best model due to the small dataset size, clearly highlighting the limitations of small samples on model generalization capabilities; Although Tanriver et al. [14] achieved an accuracy of 87.7%, the simplicity of their classification limited its clinical utility. While Lin et al. [32] and Mira et al. [35] subdivided OPMDs into low-risk and high-risk subgroups, the resulting model accuracy of 84.3% indicates that there remains significant room for improvement. In comparison, our study constructed a dataset comprising 1,348 high-quality oral images sourced from smartphone capture, professional camera acquisition, and literature sources. More significantly, we designed one of the most granular classification tasks reported in the literature: distinguishing not only between Normal, OPMDs, and OCA but further subdividing OPMDs into three specific lesion types: OLK, OLP, and OSF. To address this complex task, we proposed and validated an innovative dual-stage classification model (Swin Transformer and DenseNet-169). This model achieved a high level of comprehensive performance, outperforming the results of other classification studies referenced in Table 6. It’s worth noting that the image labels in this study lack confirmation from histopathological examination. This limitation may affect the model’s generalization ability in cases where clinical and pathological findings are inconsistent. For example, the model may struggle to accurately distinguish between OPMDs that have progressed to early carcinoma and those that have not. Therefore, in subsequent studies, we will incorporate histopathologically confirmed images to validate and enhance the model’s performance. Conclusion The application of AI for identifying OPMDs and OCA through image analysis holds substantial clinical value for screening and early prevention. This study established a high-quality oral dataset of 1,348 oral mucosal images. We further proposed a dual-stage multi-classification framework, implementing a specific model based on Swin Transformer and DenseNet-169 for five-class identification. This model demonstrated superior performance compared to conventional single-stage approaches, offering a promising computer-aided solution for OCA screening and diagnosis. Moreover, to foster further research, we have publicly released the dataset, thereby expanding the available data resources for the early detection of oral cancer. Future work Although the dual-stage classification model demonstrated significant advantages in our experiments, several challenges remain for its practical application. Most notably, our dataset lacks histopathological confirmation and the model’s accuracy in detecting OCA and OLK is relatively insufficient. Other limitations include incomplete coverage of OPMDs’ subtypes, insufficient model stability, and a lack of clinical validation. To address these issues, future work will focus on constructing a multimodal dataset of histopathologically confirmed cases that integrates images and clinical text. Leveraging this dataset, a multimodal DL model can be developed, incorporating textual features to further enhance recognition performance. Additionally, we will investigate synergistic fusion mechanisms between image segmentation and classification models to improve the extraction and recognition of key lesion features. For validation, the study will establish an external independent test set in the future. The model’s clinical applicability will be rigorously validated using an external independent test set, with performance on key metrics (e.g., accuracy, recall) systematically benchmarked against diagnoses made by clinical practitioners.