Comprehensive benchmarking of deep learning architectures for multiclass histopathological classification of oral epithelial lesions Authors: Achla Bharti, Nikita Kashyap, Mala Kamboj, Kamaldeep Joshi, Sahil Hooda, Deepika Mishra, Harpreet Singh, Debnath Pal Published: 2026 ABSTRACT BackgroundAccurate histopathological differentiation of normal oral epithelium, oral epithelial dysplasia (OED) and oral squamous cell carcinoma (OSCC) is essential but remains time consuming and prone to inter-observer variability. Deep learning provides a powerful approach for automated analysis of histopathological images in digital pathology. This study comparatively evaluated multiple deep learning architectures for automated classification of histopathological images into normal oral epithelium, OED and OSCC.MethodsA curated dataset of 2363 histopathological images captured at 10x and 40× magnifications was analyzed, comprising normal epithelium (n = 1254), OED (n = 976) and OSCC (n = 102). The dataset was divided into training (80%) and testing (20%) sets. To address class imbalance and improve model robustness, data augmentation and weighted loss strategies were implemented, expanding the training dataset to 4750 images. Seven deep learning architectures- Vision Transformer (ViT), EfficientNet-B0, InceptionV3, ResNet-50, VGG16, MobileNetV3 and YOLOv11 were trained under standardized conditions using 224×224 input resolution for 100 epochs. Model performance was assessed using accuracy, precision, recall and F1-score.ResultsEfficientNet-B0 achieved the highest validation accuracy (97.48%) and the best overall test performance with macro and weighted F1-scores of 0.97. InceptionV3, YOLOv11 and ResNet-50 also demonstrated strong classification capability (F1 ≈ 0.96). Misclassifications were rare and primarily occurred at histological transition regions between normal epithelium, dysplasia and early carcinoma.ConclusionDeep learning enables highly accurate classification of oral histopathological images. EfficientNet-B0 demonstrated superior performance, supporting the potential of AI-assisted digital pathology for early detection and objective assessment of oral precancer and cancer. Introduction Microscopic evaluation of biopsy-derived oral tissue specimens plays a pivotal role in confirming the diagnosis of Oral Epithelial Dysplasia/Oral Squamous Cell Carcinoma. During this process, pathologists examine architectural and cytological features such as epithelial arrangement, cellular atypia and nuclear abnormalities to identify malignant transformation within the oral mucosa.1 Burden of oral squamous cell carcinoma and oral epithelial dysplasia Head and neck cancers are among the ten most prevalent malignancies globally, with a significant proportion arising in the oral cavity. Oral squamous cell carcinoma (OSCC) accounts for over 90% of all oral malignancies and continues to pose a substantial public health challenge due to its high morbidity and mortality.2 Chronic exposure to carcinogenic agents such as tobacco, alcohol, and betel quid leads to cumulative genetic alterations within the oral epithelium, contributing significantly to the development of oral cancer. According to the World Health Organization (WHO), oral cancer accounts for a substantial proportion of cancer-related morbidity and mortality worldwide, particularly in developing countries where these risk factors are highly prevalent.3 Despite advances in diagnostic and therapeutic strategies, the prognosis of patients with OSCC remains unsatisfactory. The overall survival rate has shown only modest improvement over the past decades and continues to range between 50 and 70%, largely due to delayed diagnosis and the progression of precursor lesions into invasive carcinoma.4 Oral epithelial dysplasia (OED) represents a spectrum of histopathological alterations associated with oral potentially malignant disorders (OPMDs) and is widely regarded as the most important predictive indicator for malignant transformation into OSCC. Accurate identification and grading of dysplastic changes in oral epithelium are therefore critical for early detection and timely clinical intervention.5 However, histopathological evaluation of oral lesions remains largely dependent on manual microscopic examination by pathologists. Although considered the gold standard, this process is often time-consuming and subject to inter- and intra-observer variability, particularly when distinguishing subtle morphological differences between normal oral epithelium, OED and OSCC.6 Role of artificial intelligence in medical image analysis Artificial intelligence (AI), particularly deep learning, has become increasingly integrated into modern healthcare and is playing an expanding role in improving diagnostic accuracy, clinical decision-making and patient management. In dentistry, AI-based systems are being widely explored for applications such as radiographic interpretation, detection of dental caries and periapical lesions, orthodontic treatment planning, implantology and other image-guided diagnostic procedures. Recent advances in deep neural networks have enabled automated analysis of complex medical and dental images through efficient extraction of clinically relevant visual features. In oral pathology, deep learning models have shown promising capability in identifying histopathological patterns associated with dysplasia and malignancy, thereby supporting the detection and classification of oral epithelial lesions. Such AI-assisted approaches may facilitate rapid and objective image interpretation, reduce observer-related variability and improve the reproducibility and efficiency of histopathological assessment. Consequently, these technologies hold substantial potential for enhancing early detection of OSCC and supporting more effective clinical management of oral potentially malignant and malignant disorders.7, 8, 9. Artificial intelligence and transfer learning in digital pathology Recent advances in artificial intelligence have significantly transformed the field of digital pathology by enabling automated analysis of histopathological images. Transfer learning has emerged as a widely adopted approach in deep learning, particularly in medical imaging applications. By leveraging pretrained models, transfer learning addresses key challenges such as the limited availability of annotated medical datasets and the high computational cost associated with training deep learning models from scratch. This approach enhances model generalizability, improves diagnostic accuracy for specialized medical tasks, reduces the number of samples required for effective training and shortens training time. In particular, Convolutional Neural Networks (CNNs) have demonstrated remarkable capability in automatically learning hierarchical image features, thereby contributing significantly to advancements in object recognition, image segmentation and AI-assisted medical diagnosis.10,11 Architectures such as ResNet-50, VGG16, and EfficientNet-B0 have achieved strong performance in medical image classification by learning hierarchical features from large datasets. In addition to conventional CNN-based classifiers, object detection frameworks such as YOLOv11 (You Only Look Once) have gained attention for their ability to perform rapid and accurate detection of pathological features within images by simultaneously localizing and classifying regions of interest.12,13 Compared with several other deep learning frameworks, YOLO represents a highly efficient, accessible and clinically adaptable architecture for dental image analysis.14 More recently, transformer-based architectures such as the Vision Transformer have further expanded the capabilities of deep learning by capturing global contextual relationships within images. The integration of these deep learning approaches into digital pathology offers significant potential for developing computer-assisted diagnostic systems that can assist pathologists in the accurate identification and classification of oral epithelial abnormalities.15 Deep learning applications in oral histopathology In the context of oral pathology, the application of deep learning to histopathological images has gained increasing attention for its potential to support pathologists in diagnostic decision-making and improve the efficiency and accuracy of disease detection.12 Automated classification systems can assist in identifying subtle architectural and cytological alterations that characterize dysplasia and malignant changes, thereby facilitating objective and reproducible diagnosis. Nevertheless, comparative evaluations of different deep learning architectures for distinguishing normal oral epithelium, OED and OSCC remain relatively limited, particularly in datasets representing the histological continuum of oral carcinogenesis.16, 17,18 Research gap and rationale for benchmarking multiple architectures Despite considerable advancements in histopathological imaging and artificial intelligence based diagnostic systems, accurate differentiation between normal oral epithelium, oral epithelial dysplasia and OSCC remains a significant challenge in routine pathological assessment. Existing diagnostic approaches are still largely dependent on manual interpretation, which may be affected by observer variability, subjective assessment and difficulty in recognizing subtle morphological transitions across the spectrum of oral carcinogenesis. Furthermore, many previously reported AI-based studies have focused primarily on binary classification tasks or individual deep learning models, limiting comprehensive understanding of the comparative diagnostic capabilities of different architectures, thereby providing insufficient evidence regarding the optimal architecture for multiclass discrimination of oral epithelial lesions across varying histological complexities. Additionally, there remains a lack of standardized benchmarking studies integrating both CNN-based image classification networks and transformer/object-detection frameworks within the same experimental setting using oral histopathological datasets. To address this gap, the present study incorporates multiple state-of-the-art architectures representing different computational paradigms and design philosophies, including conventional deep CNNs (VGG16, ResNet-50, InceptionV3), lightweight and computationally efficient models suitable for resource-constrained deployment (MobileNetV3, EfficientNet-B0), transformer-based architecture capable of capturing global contextual relationships (Vision Transformer; ViT) and a real-time object detection framework with enhanced localization capability (YOLOv11). The inclusion of these architectures was intended to provide a comprehensive benchmarking framework encompassing variations in network depth, feature extraction strategy, parameter efficiency, receptive field learning and spatial localization ability, all of which may influence diagnostic performance in histopathological interpretation. Furthermore, by integrating advanced preprocessing techniques, augmentation strategies and optimized training frameworks, the study seeks to improve automated discrimination between normal oral epithelium OED and OSCC. The use of systematically curated histopathological datasets and standardized benchmarking conditions further contributes toward establishing a reliable baseline framework for future AI-assisted oral pathology research and clinical applications. Objectives of the study Accordingly, the present work comparatively evaluates a diverse spectrum of deep learning architectures for automated histopathological classification of normal oral epithelium, OED and OSCC. Using a curated dataset of histopathological images acquired at different magnifications, these architectures were systematically trained and evaluated for multiclass diagnostic performance. By benchmarking models with fundamentally different learning mechanisms, the study seeks to identify the most effective and generalizable architecture for accurate histopathological discrimination of oral epithelial lesions, while also exploring the translational potential of AI-assisted digital pathology for early detection and diagnosis of oral precancerous and malignant conditions. Materials and methods Study design The present retrospective study was conducted in the Department of Oral Pathology and Microbiology at the Post Graduate Institute of Dental Sciences, Rohtak, Haryana, as part of a national research initiative funded by the Indian Council of Medical Research (ICMR). The study utilized hematoxylin and eosin (H&E) stained histopathological slides of oral mucosal biopsy specimens retrieved from the departmental archives. The dataset comprised histopathological confirmed cases representing three diagnostic categories: normal oral epithelium, OED and OSCC. The diagnosis of these lesions was established based on the histopathological criteria outlined by the World Health Organization,19 which considers characteristic cytological and architectural alterations observed in the oral epithelium. These digitized histopathological images were subsequently used for the development and comparative evaluation of deep learning models for automated classification of the three diagnostic categories. Prior to data collection, ethical approval was obtained from the Institutional Review Board of the institute (IRB No.: PGIDS/BHRC/23/15). All patient identifiers were removed and the dataset was anonymized to ensure confidentiality and compliance with ethical research standards. Data acquisition process Representative microscopic fields from each slide were visualized and digitally captured using an Olympus BX53 microscope equipped with a high-resolution digital camera (Model DP23). Images were acquired at two magnification levels (10× and 40x) to capture both overall epithelial architecture and detailed cytological features relevant for the classification of oral epithelial lesions. All imaging parameters, including illumination, white balance and exposure settings, were standardized during image acquisition to minimize inter-sample variability and ensure consistency within the dataset. All histopathological slides were examined by three experienced oral pathologists to confirm diagnostic adequacy and representative epithelial morphology. Areas affected by poor staining, sectioning artifacts, tissue folds or inadequate epithelial representation were excluded from further analysis. In cases of disagreement, the slides were re-evaluated collaboratively and a final diagnosis was established by consensus to ensure reliable ground-truth labeling for the deep learning models. Dataset overview The histopathological image dataset comprised of 2363 H&E-stained digital images categorized into three diagnostic classes: normal oral mucosa, OED and OSCC. Images were captured at two magnification levels (10× and 40x) to ensure representation of both epithelial architectural patterns and cytological characteristics relevant for lesion classification. As summarized in Table 1, the dataset included 102 images of OSCC (26 at 10x and 76 at 40x), 976 images of OED (327 at 10x and 649 at 40x) and 1254 images of normal oral mucosa (193 at 10x and 786 at 40x). Overall, the dataset consisted of 546 images captured at 10× magnification and 1817 images at 40× magnification. The distribution of images revealed a notable class imbalance, which is commonly encountered in medical imaging datasets. The OSCC category contained substantially fewer images compared with the dysplasia and normal classes. Such imbalance can influence model learning and potentially bias predictions toward the majority classes. Therefore, careful dataset partitioning and preprocessing strategies were implemented to ensure reliable model development and evaluation.Table 1Distribution of the initial image dataset across diagnostic classes and magnification levels.Table 1ClassTotal Images10× Magnifications40× MagnificationCancer1022676Dysplasia976327955Normal1254193786Total23635461817 Table 1Distribution of the initial image dataset across diagnostic classes and magnification levels.Table 1ClassTotal Images10× Magnifications40× MagnificationCancer1022676Dysplasia976327955Normal1254193786Total23635461817 Distribution of the initial image dataset across diagnostic classes and magnification levels. Data partitioning For model development and evaluation, the dataset was divided into training and testing subsets using an 80:20 split. This resulted in 1887 images allocated for training and 476 images reserved for independent testing. To ensure unbiased model evaluation, the dataset was divided into training and testing subsets before any data augmentation procedures were applied, thereby preventing information leakage between the two sets. The resulting distribution included 80 cancer images for training and 22 for testing, 1025 dysplasia images for training and 257 for testing and 782 normal images for training and 197 for testing, as presented in Table 2. The training dataset was used to develop and optimize the deep learning models, whereas the testing dataset was used exclusively for final performance evaluation to assess model generalizability.Table 2Pre-augmentation split of the dataset into training (80%) and testing cohorts (20%).Table 2ClassTotal ImagesTrainTestCancer1028022Dysplasia12821025257Normal979782197Total23631887476 Table 2Pre-augmentation split of the dataset into training (80%) and testing cohorts (20%).Table 2ClassTotal ImagesTrainTestCancer1028022Dysplasia12821025257Normal979782197Total23631887476 Pre-augmentation split of the dataset into training (80%) and testing cohorts (20%). Before model training, all images underwent preprocessing to ensure compatibility with the deep learning architectures used in the study. Specifically, the images were resized to a uniform dimension of 224 × 224 pixels, which is a standard input size for several convolutional neural network models. This preprocessing step ensured consistent input formatting while preserving the essential morphological features required for automated histopathological classification. Image preprocessing and patch extraction The acquired histopathological images were first subjected to a series of preprocessing steps prior to deep learning model development to improve image quality and ensure uniformity across the dataset. Initially, all images were carefully inspected to remove blurred, poorly focused or artifact-containing regions. The selected images were then resized and formatted to maintain a consistent input dimension compatible with the deep learning architectures used in the study. To facilitate efficient training of the models and to capture localized histopathological features, the digitized images were divided into smaller non-overlapping image patches. Patch extraction was performed from representative epithelial regions while avoiding areas with excessive connective tissue, background or staining artifacts. This approach enabled the generation of multiple training samples from a single image while preserving diagnostically relevant morphological patterns. Images/patches derived from the same histopathological slide were restricted to only one subset to prevent information leakage. Stain normalization Despite the use of a standardized H&E staining protocol, slight variations in staining intensity and color tone were observed across the digitized histopathological images. To minimize these variations and ensure color consistency throughout the dataset, stain normalization was performed using the Reinhard normalization method. This approach adjusts the color profile of the input image to correspond with that of a predefined reference image (Fig. 1).Fig. 1Illustration of stain normalization: (A) original image with variability in H&E staining intensity and color; (B) normalized image with standardized color balance. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.) Illustration of stain normalization: (A) original image with variability in H&E staining intensity and color; (B) normalized image with standardized color balance. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.) The Reinhard technique standardizes the color distribution of an input image by transforming it to match the statistical color characteristics of a selected reference image. Initially, the input image is converted from the RGB color space to the LAB color space, which separates image information into three channels representing lightness (L) and chromatic components along the green-red (A) and blue-yellow (B) axis. This color space is particularly suitable for normalization because it better represents perceptual color differences. Following this transformation, the mean and standard deviation of each LAB channel of the input image are adjusted to match those of the reference image. By aligning the statistical properties of the color channels, the Reinhard method effectively harmonizes staining appearance across images while preserving the underlying histopathological structures. Through this normalization process, variability arising from staining procedures, tissue preparation and image acquisition is reduced. Consequently, the deep learning models are able to focus on relevant morphological and structural features rather than variations in color intensity or staining artifacts. The general form of the stain normalization equation is:In =Io ∗ (1+k1 ∗ (Lo −μL) + k2 ∗ (So −μS))where: In = normalized output image. Io = original image k1 and k2 = tuning constants that are chosen to optimize the appearance of the normalized image Lo, So = mean brightness and saturation of the original image μL and μS = corresponding statistics of the reference image. Experimental configuration This study conducted a comparative evaluation of several deep learning architectures including Vision Transformer (ViT), EfficientNet-B0, InceptionV3, ResNet-50, VGG16, MobileNetV3 and YOLOv11 for classifying histopathological images of normal oral epithelium, OED and OSCC, with the goal of identifying the most effective model. The computational workflow was implemented in a Jupyter Notebook environment using Python and the PyTorch deep learning framework. The histopathological image dataset was organized within structured training and testing directories and uploaded to the computational workspace for model development and evaluation. All experiments were conducted using an NVIDIA GeForce RTX 4060 Laptop GPU with 8 GB memory acceleration, enabling efficient parallel computation during model training. The models were initialized using pretrained ImageNet weights to facilitate transfer learning and subsequently fine-tuned using the oral histopathological dataset. The training pipeline involved sequential preprocessing, augmentation, batch loading, model optimization and validation monitoring. The AdamW optimizer was used with categorical cross-entropy loss and class weighting strategies to address class imbalance. Early stopping with a patience value of 30 epochs was implemented to prevent overfitting and improve computational efficiency. The complete workflow ensured standardized training conditions across all evaluated architectures to enable fair comparative benchmarking. Deep learning model development Model architectures The deep learning architectures evaluated in the study included:1.Vision Transformer (ViT): The Vision Transformer architecture adapts the Transformer model, originally designed for natural language processing, to computer vision. It treats an image as a sequence of fixed-size patches and uses self-attention mechanisms to weigh the importance of all patches relative to each other.202.EfficientNet-B0: A state-of-the-art architecture known for its efficiency and performance. Its core building block is the Mobile inverted Bottleneck Convolution (MBConv), enhanced with squeeze-and-excitation blocks, improving feature representation with minimal performance overhead.213.Inception-v3: A deep convolutional neural network developed by Google. ‘Inception modules’ perform convolutions with multiple kernel sizes in parallel, capturing and aggregating visual features at different scales efficiently.224.ResNet-50: A foundational architecture with the concept of residual learning through ‘skip connections’, allowing gradients to bypass layers during backpropagation and mitigating the vanishing gradient problem effectively.235.VGG16: A classic deep CNN known for its architectural simplicity and uniformity. It consists of sequential stacks of 3x3 convolutional layers, demonstrating that significant depth is critical for achieving high performance in image recognition tasks.246.Mobilenet-v3: A lightweight architecture designed for deployment on mobile and edge devices. Its efficiency stems from the use of depth-wise separable convolutions, which significantly reduce the number of parameters and computations.257.YOLOv11: YOLOv11, being the latest iteration in the You Only Look Once series, is known for its exceptional balance between speed and accuracy in real-time object detection and image classification.26 1.Vision Transformer (ViT): The Vision Transformer architecture adapts the Transformer model, originally designed for natural language processing, to computer vision. It treats an image as a sequence of fixed-size patches and uses self-attention mechanisms to weigh the importance of all patches relative to each other.20 Vision Transformer (ViT): The Vision Transformer architecture adapts the Transformer model, originally designed for natural language processing, to computer vision. It treats an image as a sequence of fixed-size patches and uses self-attention mechanisms to weigh the importance of all patches relative to each other.20 2.EfficientNet-B0: A state-of-the-art architecture known for its efficiency and performance. Its core building block is the Mobile inverted Bottleneck Convolution (MBConv), enhanced with squeeze-and-excitation blocks, improving feature representation with minimal performance overhead.21 EfficientNet-B0: A state-of-the-art architecture known for its efficiency and performance. Its core building block is the Mobile inverted Bottleneck Convolution (MBConv), enhanced with squeeze-and-excitation blocks, improving feature representation with minimal performance overhead.21 3.Inception-v3: A deep convolutional neural network developed by Google. ‘Inception modules’ perform convolutions with multiple kernel sizes in parallel, capturing and aggregating visual features at different scales efficiently.22 Inception-v3: A deep convolutional neural network developed by Google. ‘Inception modules’ perform convolutions with multiple kernel sizes in parallel, capturing and aggregating visual features at different scales efficiently.22 4.ResNet-50: A foundational architecture with the concept of residual learning through ‘skip connections’, allowing gradients to bypass layers during backpropagation and mitigating the vanishing gradient problem effectively.23 ResNet-50: A foundational architecture with the concept of residual learning through ‘skip connections’, allowing gradients to bypass layers during backpropagation and mitigating the vanishing gradient problem effectively.23 5.VGG16: A classic deep CNN known for its architectural simplicity and uniformity. It consists of sequential stacks of 3x3 convolutional layers, demonstrating that significant depth is critical for achieving high performance in image recognition tasks.24 VGG16: A classic deep CNN known for its architectural simplicity and uniformity. It consists of sequential stacks of 3x3 convolutional layers, demonstrating that significant depth is critical for achieving high performance in image recognition tasks.24 6.Mobilenet-v3: A lightweight architecture designed for deployment on mobile and edge devices. Its efficiency stems from the use of depth-wise separable convolutions, which significantly reduce the number of parameters and computations.25 Mobilenet-v3: A lightweight architecture designed for deployment on mobile and edge devices. Its efficiency stems from the use of depth-wise separable convolutions, which significantly reduce the number of parameters and computations.25 7.YOLOv11: YOLOv11, being the latest iteration in the You Only Look Once series, is known for its exceptional balance between speed and accuracy in real-time object detection and image classification.26 YOLOv11: YOLOv11, being the latest iteration in the You Only Look Once series, is known for its exceptional balance between speed and accuracy in real-time object detection and image classification.26 Handling class imbalance Class imbalance was addressed using a dual approach: by weighting the learning algorithm to prioritize underrepresented classes and by augmenting the training dataset to improve class distribution.i)Weighted Loss Function i)Weighted Loss Function Weighted Loss Function The first approach used a weighted loss function, where each class was assigned a weight inversely proportional to its frequency in the training dataset. The weight for each class was calculated using the following formula:Weight=TotalSamplesNumberofClasses×SamplesinClass This weighting strategy imposes a higher penalty on the model for misclassifying images belonging to minority classes than for errors involving majority-class samples. Consequently, the learning algorithm is encouraged to focus on capturing distinctive features of underrepresented yet clinically significant classes, such as dysplasia and carcinoma, rather than being biased toward the majority class.ii)Data augmentation ii)Data augmentation Data augmentation To further address the class imbalance and enhance the diversity of the training data, data augmentation techniques were applied to the images in the training set. The augmentation pipeline was implemented on the 1887 training images, expanding the dataset to a balanced set of 4750 images. Following augmentation, the training dataset consisted of 1350 cancer images, 1550 dysplasia images and 1850 normal images, thereby improving class representation during model training. The augmentation strategy was designed to introduce realistic variations while preserving the essential structural and color characteristics of histopathological images. Geometric transformations, including horizontal and vertical flipping and random rotations up to ±20°, were applied to simulate variations in tissue orientation and slide positioning during image acquisition. In addition, photometric transformations such as adjustments in brightness, contrast, saturation and hue were incorporated to account for potential variations in illumination and staining appearance. To further improve model robustness, Gaussian noise was added to mimic minor electronic or imaging noise that can naturally occur in digital images. These augmentation techniques generated diverse yet biologically meaningful training samples, enabling the deep learning models to learn invariant features associated with the histopathological patterns of oral lesions. The testing dataset was intentionally kept unchanged to ensure an unbiased evaluation of model performance on unseen data. The final distribution of images following data augmentation is summarized in Table 3.Table 3Class-wise distribution of the augmented training set and the original testing set.Table 3Total ImagesTrainTestCancer1372135022Dysplasia17471550257Normal21071850197Total52264750476 Table 3Class-wise distribution of the augmented training set and the original testing set.Table 3Total ImagesTrainTestCancer1372135022Dysplasia17471550257Normal21071850197Total52264750476 Class-wise distribution of the augmented training set and the original testing set. Evaluation metrics To assess the classification performance of the models, several evaluation metrics were employed, including accuracy, precision, recall and F1-score. Accuracy represents the overall proportion of correctly classified instances among all samples. Precision measures the reliability of the model's positive predictions by calculating the proportion of true positives among all predicted positives. Recall evaluates the model's ability to correctly identify relevant instances by determining the proportion of true positives detected out of all actual positive cases. The F1-score, defined as the harmonic mean of precision and recall, provides a balanced measure that accounts for both false positive and false negative errors. The statistical analysis for the image classification models was performed using the following computational approaches:Accuracy=TP+TNTP+FP+TN+FN∗100Precision=TPTP+FP∗100Recall=TPTP+FN∗100F1−score=2∗Precision∗Recall1∗Precision+Recall∗100•True positive (TP): positive outcomes that the model predicted correctly.•False positive (FP): positive outcomes that the model predicted incorrectly.•True negative (TN): negative outcomes that the model predicted correctly.•False negative (FN): negative outcomes that the model predicted incorrectly. •True positive (TP): positive outcomes that the model predicted correctly. True positive (TP): positive outcomes that the model predicted correctly. •False positive (FP): positive outcomes that the model predicted incorrectly. False positive (FP): positive outcomes that the model predicted incorrectly. •True negative (TN): negative outcomes that the model predicted correctly. True negative (TN): negative outcomes that the model predicted correctly. •False negative (FN): negative outcomes that the model predicted incorrectly. False negative (FN): negative outcomes that the model predicted incorrectly. Results Quantitative analysis of model performance Performance evaluation on the unseen test set (n = 476) was conducted using macro and weighted averages of precision, recall and F1-score as presented in Table 4. Seven deep learning architectures MobileNetV3, VGG16, Vision Transformer (ViT), ResNet-50, YOLOv11, Inception-V3 and EfficientNet-B0 were systematically assessed for multi-class classification of normal oral epithelium, OED and OSCC. MobileNetV3 demonstrated comparatively lower macro-averaged performance with precision, recall and F1-score of 0.92, 0.81 and 0.85 respectively, while the corresponding weighted values were all 0.89. VGG16 showed improved and stable classification performance, achieving macro-averaged precision, recall and F1-score of 0.92, 0.89 and 0.90 respectively, with corresponding weighted values were 0.90 for all three metrics. Vision Transformer (ViT) further demonstrated strong performance, with macro-averaged precision, recall and F1-score of 0.93, 0.95 and 0.94 and weighted averages of 0.93 across all metrics, indicating effective classification capability across the classes.Table 4Comparative performance analysis of deep learning models on the test dataset.Table 4Model NameMacro AverageWeighted AveragePrecisionRecallF1-ScorePrecisionRecallF1-ScoreMobilenet-v30.920.810.850.890.890.89VGG160.920.890.900.900.900.90Vision Transformer (ViT)0.930.950.940.930.930.93Resnet-500.960.960.960.960.960.96YOLOv110.960.960.960.960.960.96Inception-v30.970.980.970.960.960.96EfficientNet-B00.980.970.970.980.970.97 Table 4Comparative performance analysis of deep learning models on the test dataset.Table 4Model NameMacro AverageWeighted AveragePrecisionRecallF1-ScorePrecisionRecallF1-ScoreMobilenet-v30.920.810.850.890.890.89VGG160.920.890.900.900.900.90Vision Transformer (ViT)0.930.950.940.930.930.93Resnet-500.960.960.960.960.960.96YOLOv110.960.960.960.960.960.96Inception-v30.970.980.970.960.960.96EfficientNet-B00.980.970.970.980.970.97 Comparative performance analysis of deep learning models on the test dataset. Among the higher-performing models, ResNet-50 and YOLOv11 demonstrated excellent and consistent performance, each achieving macro-averaged precision, recall and F1-score of 0.96, with corresponding weighted values also reaching 0.96. Inception-V3 showed slightly improved performance, attaining macro-averaged precision, recall and F1-score of 0.97, 0.98 and 0.97 respectively, while maintaining weighted values of 0.96 across the evaluation metrics. Notably, EfficientNet-B0 achieved the best overall classification performance, with macro-averaged precision, recall and F1-score of 0.98, 0.97 and 0.97 respectively with identical weighted values, reflecting highly balanced and robust classification across all categories. Overall, the results demonstrate that advanced CNN architectures particularly EfficientNet-B0, Inception-V3, ResNet-50 and YOLOv11 provide superior and consistent performance for histopathological classification of normal oral epithelium, OED and OSCC, highlighting their potential for integration into AI-assisted digital oral pathology workflows. Training and validation behavior The training and validation curves of all evaluated models illustrated distinct learning dynamics and generalization capabilities, with accuracy and loss plotted as functions of training epochs to assess their performance trends. Fig. 2 presents the training and validation performance curves for each deep learning architecture across the training process, with all models trained under identical experimental conditions on an NVIDIA RTX 4060 GPU (8 GB) for 100 epochs using an input resolution of 224 × 224 pixels.Fig. 2Training and validation accuracy and loss curves across epochs for MobileNetV3, VGG-16, ViT, ResNet-50, YOLOv11, InceptionV3, and EfficientNet-B0. Training and validation accuracy and loss curves across epochs for MobileNetV3, VGG-16, ViT, ResNet-50, YOLOv11, InceptionV3, and EfficientNet-B0. MobileNetV3 showed consistently high training accuracy; however, its validation accuracy plateaued at approximately 89%, accompanied by noticeable fluctuations in validation loss, suggesting instability in handling unseen data. VGG-16 exhibited a steady increase in both training and validation accuracy, reaching around 90% validation accuracy, although a persistent gap between the curves and comparatively higher validation loss indicated mild overfitting. In comparison, the Vision Transformer (ViT) demonstrated strong learning behavior, with validation accuracy exceeding 93%, smooth convergence of loss curves, and a minimal gap between training and validation performance, reflecting good generalization. Further improvement was observed with ResNet-50, which rapidly achieved validation accuracy of approximately 96%, with closely aligned training and validation curves, indicating minimal overfitting and robust generalization. A similar trend was noted for the YOLOv11 classification model, where validation accuracy reached nearly 96% alongside stable and consistent reductions in both training and validation loss, highlighting effective learning. The InceptionV3 model also performed strongly, achieving validation accuracy above 96%; however, a slight divergence between training and validation loss in later epochs suggested the early onset of overfitting. Among all models, EfficientNet-B0 demonstrated the most optimal behavior, with validation accuracy exceeding 97% and closely mirroring training accuracy, while the loss curves converged smoothly at low values, confirming superior stability, robustness and generalization performance. Comparative validation accuracies (Fig. 3) further highlighted inter-model performance differences, with EfficientNet-B0 achieving the highest validation accuracy (97.48%), followed by InceptionV3 (96.43%), YOLOv11 (96.21%), and ResNet-50 (95.79%). ViT demonstrated moderate performance with a validation accuracy of 93.07%, whereas comparatively lower accuracies were observed for the lightweight models, VGG16 (89.71%) and MobileNetV3 (88.66%) indicating relatively reduced generalization capability on the validation dataset. This analysis demonstrated the superior capability of the EfficientNet-B0 architecture for histopathological image classification.Fig. 3Comparison of validation accuracy across all evaluated deep learning models. Comparison of validation accuracy across all evaluated deep learning models. Comprehensive performance analysis of EfficientNet-B0 EfficientNet-B0 was further examined through a class-wise performance analysis to better understand its diagnostic strengths and limitations across the three histopathological categories. The detailed performance metrics for each class are presented in Table 5.Table 5Class-wise performance metrics of the EfficientNet-B0 model on the test dataset.Table 5ClassPrecisionRecallF1-ScoreCancer1.000.950.98Dysplasia0.961.000.98Normal0.990.940.97Macro Average0.980.970.97Weighted Average0.980.970.97 Table 5Class-wise performance metrics of the EfficientNet-B0 model on the test dataset.Table 5ClassPrecisionRecallF1-ScoreCancer1.000.950.98Dysplasia0.961.000.98Normal0.990.940.97Macro Average0.980.970.97Weighted Average0.980.970.97 Class-wise performance metrics of the EfficientNet-B0 model on the test dataset. For the cancer class, the model achieved a precision of 1.00, recall of 0.95 and an F1-score of 0.98 indicating the absence of false-positive predictions with only a small proportion of missed cancer cases. In the dysplasia category, EfficientNet-B0 demonstrated perfect recall (1.00) with a precision of 0.96 and an F1-score of 0.98, confirming complete detection of dysplastic lesions with minimal false positives. Similarly, for the normal epithelium class, the model obtained a precision of 0.99, recall of 0.94, and an F1-score of 0.97, reflecting high reliability in identifying non-pathological tissue with only occasional misclassification with dysplastic images. Overall, the model achieved macro-average precision, recall and F1-score of 0.98, 0.97 and 0.97 respectively, with weighted metrics reflecting similar values, demonstrating highly balanced and robust classification performance across the three diagnostic classes. Error analysis An analysis of the misclassification patterns infers that the errors were not randomly distributed but predominantly occurred at the boundaries between closely related diagnostic categories, particularly between dysplasia and early OSCC. This likely reflects the histopathological overlap, progressive biological continuum of oral epithelial transformation and the inherent diagnostic ambiguity encountered even among expert oral pathologists. These findings suggest that the model learned clinically meaningful histological representations rather than dataset-specific artifacts. The recall value of 0.95 for the cancer class suggests that a single cancer case (approximately 5% of the 22 cancer images in the test set) was misclassified. Similarly, the recall of 0.94 for the normal class indicates that a small proportion of normal samples were incorrectly predicted as dysplasia. This pattern of boundary-level misclassification is further illustrated by the model's confusion matrix (Fig. 4), which highlights the limited overlap between adjacent histopathological classes. This pattern of boundary-level misclassification is consistent with the diagnostic challenges and inherent observer variability encountered in human histopathological evaluation.Fig. 4Confusion matrix illustrating the classification performance of EfficientNet-B0 on the unseen test dataset. Confusion matrix illustrating the classification performance of EfficientNet-B0 on the unseen test dataset. Discussion Accurate histopathological differentiation between normal oral epithelium, oral epithelial dysplasia and oral squamous cell carcinoma is essential for determining disease progression risk and guiding appropriate clinical management. However, the interpretation of microscopic features in oral epithelial lesions is often complex and subject to inter- and intra-observer variability among pathologists, particularly in borderline cases such as early dysplastic changes. These challenges highlight the need for objective and reproducible diagnostic support systems. In this context, the application of deep learning based computational models offers a promising strategy to assist in the automated classification of oral histopathological images, potentially improving diagnostic consistency and efficiency in digital oral pathology.27, 28, 29. A major research gap in the field of AI-assisted oral histopathology is the limited availability of systematic comparative benchmarking studies evaluating multiple deep learning architectures for multiclass classification of normal oral epithelium, OED and OSCC. Most existing studies have primarily focused on binary classification tasks, single-model evaluation or limited datasets, with comparatively less emphasis on the histopathological continuum between normal epithelium, dysplasia and early carcinoma. Furthermore, limited studies have comparatively assessed convolutional neural networks, transformer-based architectures and YOLO-based frameworks under standardized experimental conditions to determine their relative effectiveness, generalization capability and clinical applicability in oral histopathological image analysis. Therefore, the present study conducted a systematic evaluation of seven state-of-the-art deep learning architectures for the automated histopathological classification of oral epithelial lesions. Among the evaluated models, EfficientNet-B0 emerged as the best-performing architecture, achieving a validation accuracy of 97.48% and an F1-score of 0.97, highlighting its effective balance between computational efficiency and strong feature representation capability. Dataset characteristics and class imbalance The superior performance of EfficientNet-B0, along with the variability observed across other architectures, can be better understood in the context of the dataset characteristics and the strategies employed to address them. The dataset exhibited a pronounced class imbalance, a common challenge in medical image analysis. The ‘Cancer’ class, comprising only 102 images, was markedly underrepresented compared to the ‘Dysplasia’ and ‘Normal’ classes, which included 976 and 1254 images, respectively. Such imbalance poses a significant risk of model bias, potentially leading to misclassification of critical cases for example, the model might incorrectly predict all samples as non-cancerous, thereby failing to detect the most clinically significant lesions. Addressing this issue requires the implementation of robust data handling strategies, including balanced training approaches, targeted augmentation and careful evaluation protocols to ensure reliable model performance across all classes. Effectiveness of data balancing and augmentation The application of weighted loss optimization combined with controlled image augmentation effectively addressed inherent class imbalance of the dataset. This approach ensured that minority class features particularly those of cancerous tissues were effectively represented during training. Moreover, the augmentation procedures preserved critical histological textures and color fidelity, likely enhancing feature extraction and contributing to the improved performance of the convolutional architectures. Model comparison and architectural insights A comparative analysis of the evaluated architectures provides important insights into their relative performance and suitability for histopathological image classification. EfficientNet-B0 achieved the highest performance, largely attributable to its compound scaling strategy, which optimally balances network depth, width and input resolution. This design facilitates effective hierarchical feature learning, enabling the model to capture both fine-grained cellular details and broader tissue-level patterns. InceptionV3 and YOLOv11 exhibited competitive performance, reflecting the advantages of architectures capable of multi-scale feature fusion, which allows simultaneous extraction of features at different spatial resolutions a key requirement for heterogeneous histopathological images where lesions vary in size and morphology. In contrast, MobileNetV3 and VGG16 showed comparatively lower classification accuracy. Their limited representational capacity and shallower feature hierarchies likely constrained their ability to capture complex morphological variations, making them more susceptible to overfitting, especially on underrepresented classes such as cancer. The Vision Transformer (ViT) displayed strong generalization and stable learning; however, its performance lagged behind convolutional models. The patch-based tokenization approach in ViT may result in the loss of fine-grained morphological cues that are crucial for distinguishing subtle epithelial changes, particularly in dysplastic or early carcinoma lesions. While ViT demonstrates promise for large-scale image recognition, its current implementation appears less suited for the nuanced texture- and structure-dependent features inherent in oral histopathology. Overall, these observations suggest that convolutional architectures with hierarchical and multi-scale feature extraction capabilities are particularly effective for automated classification of oral epithelial lesions, while transformer-based models may require adaptations, such as hybrid CNN-transformer frameworks or higher-resolution patch strategies, to fully leverage their potential in this domain. Clinical interpretability and relevance The precision-recall dynamics of EfficientNet-B0 revealed clinically meaningful behavior, indicating that the model is well suited for potential diagnostic support in oral histopathology. Notably, the model achieved a precision of 1.00 for the cancer class, indicating the absence of false-positive predictions. Each instance in which the model predicted cancer corresponded to a true cancer case in the ground truth. This characteristic is particularly important in clinical workflows, as it reduces the likelihood of unnecessary patient anxiety, additional diagnostic procedures, or unwarranted interventions triggered by false alarms. Such performance highlights the model's reliability and strengthens its potential utility as a supportive tool in the histopathological assessment of oral epithelial lesions. Similarly, the perfect recall of 1.00 for the dysplasia class ensures that no potentially premalignant lesion is overlooked, an essential requirement for screening-level applications. Such high sensitivity is particularly valuable in screening contexts, where the primary objective is the early identification of all potential precancerous lesions for further clinical evaluation. The ability of the model to maintain high specificity in cancer detection while simultaneously achieving maximal sensitivity for dysplasia indicating that the model has learned clinically meaningful and robust decision boundaries. Misclassification analysis indicated that the errors were not randomly distributed but were primarily concentrated at the boundaries between adjacent diagnostic categories. Such patterns suggest that the model encountered difficulty mainly in borderline cases where morphological features overlap between classes. This tendency is consistent with the continuum of epithelial transformation, where distinctions between normal epithelium, dysplasia and early carcinoma may be subtle. The confusion matrix further supports this observation, demonstrating that most misclassifications occurred between neighboring categories rather than between biologically distinct classes. The error analysis reflected pathologist-level diagnostic challenges, with misclassifications predominantly occurring along the biological continuum between normal, dysplastic and malignant states. This indicates that the model has learned clinically relevant morphological patterns rather than spurious associations, supporting its potential applicability in clinical settings. Comparative performance with existing literature Recent studies have increasingly explored deep learning approaches for automated histopathological classification of oral epithelial lesions, demonstrating encouraging diagnostic performance. The findings of the present study are consistent with these developments and further reinforce the applicability of advanced convolutional architectures for multi-class classification of oral epithelial lesions. Maia et al. (2024) evaluated several deep learning and few-shot learning approaches, including DenseNet-121, transformer-based networks and metric-learning frameworks such as Siamese Network, Triplet Network and ProtoNet, using 3763 histopathological patches from the P-NDB-UFES dataset. Their study focused on the classification of non-dysplastic epithelium, dysplasia and OSCC. Among the evaluated architectures, DenseNet-121 achieved the best performance, with a balanced accuracy of 91.91% and precision and recall values of 91.93%.30 Although this study demonstrated the feasibility and effectiveness of deep learning for oral lesion classification, the performance observed in the present study was higher, with EfficientNet-B0 achieving an accuracy of 97.48% and an F1-score of 0.97, indicating improved discriminative capability under standardized training and optimization strategies. Building on these developments, Tafala et al. (2025) proposed DeepPatchNet, a hybrid architecture integrating DeepLabV3+ and ConvMixer with Grad-CAM visualization to enhance model interpretability. The model was trained using the NDB-UFES dataset (3763 images) along with an additional Oral Cavity-Derived Cancer dataset containing 1020 images, achieving an overall accuracy of 86.71%, with precision, recall and F1-score values of approximately 86.8%.31 Compared with these findings, the models evaluated in the present study, particularly EfficientNet-B0, demonstrated substantially higher classification performance, highlighting the advantages of optimized architecture scaling and training strategies for histopathological image analysis. In contrast to end-to-end deep learning approaches, Mandal et al. (2025) adopted a hybrid framework combining deep feature extraction using ResNet-50 with conventional machine learning classifiers, including KNN, linear SVM, decision tree, random forest, logistic regression and naïve Bayes. Their methodology also incorporated multimodal integration of histopathological features with clinical and demographic data. Using 237 samples from the NDB-UFES dataset to classify OSCC, leukoplakia with dysplasia and leukoplakia without dysplasia, the KNN classifier achieved the highest accuracy of 89%, while the linear SVM achieved an accuracy of 81%.32 Although the integration of clinical information demonstrated potential benefits, the overall classification performance remained lower than that achieved by fully deep learning-based models such as those evaluated in the present study. More recently, Akram et al. (2026) developed a Deep Visual Detection System incorporating EfficientNetB3, DenseNet121 and ResNet50 architectures. The model was trained on two datasets, including a Kaggle dataset containing 5192 images for binary classification of normal epithelium and OSCC and the NDB-UFES dataset with 3763 images for multi-class classification of OSCC, leukoplakia with dysplasia and leukoplakia without dysplasia. Among the evaluated architectures, EfficientNetB3 achieved the best performance, with an accuracy of 97.16% and corresponding precision, recall and F1-score values of 97.16% in the multi-class classification task.33 These findings are comparable with the results of the present study, where EfficientNet-B0 achieved a slightly higher accuracy of 97.48%, with precision, recall and F1-score values of approximately 0.98, 0.97 and 0.97 respectively, for multi-class classification of normal epithelium, OED and OSCC. The key characteristics and performance outcomes of these studies are comparatively summarized in Table 6.Table 6Comparative analysis of previously published deep learning architectures for histopathological classification of normal epithelium, OED and OSCC.Table 6Authors, YearCountryAI architectureDL/ML/HybridImage countClassification approachResultsMaia et al. 202430BrazilCNNs (DenseNet-121), Transformer neural networks & Few-shot learning models (Siamese Network, Triplet Network, ProtoNet)DL3763 patches (P-NDB-UFES datasetNo Dysplasia vs Dysplasia vs OSCCBest Performing Model DenseNet-121Balanced Accuracy (BCC)- 91.91%Precision- 91.93%Recall- 91.93%Tafala et al. 202531MoroccoDeepPatchNet (DeepLabV3+ + ConvMixer) + Grad-CAMDL3763 (NDB-UFES dataset) 1020 [Oral Cavity-Derived Cancer (OCDC) dataset]No Dysplasia vs Dysplasia vs OSCCAccuracy- 86.71%Precision- 86.80%Recall- 86.71%F1-score- 86.75%Mandal et al. 202532IndiaResNet50 + Machine Learning classifiers (KNN, Linear SVM, Decision Tree, Random Forest, Logistic Regression, Naïve Bayes); Multimodal integration (histopathological image features + clinical-demographic data)Hybrid237 (NDB-UFES dataset)OSCC vs leukoplakia with dysplasia vs leukoplakia without dysplasiaKNN:Highest accuracy- 89%Linear SVM: Accuracy- 81%Akram et al. 202633PakistanDeep Visual Detection System (DVDS); EfficientNetB3; DenseNet121; ResNet50DL5192 (Normal vs OSCC- Kaggle)3763 (OSCC, leukoplakia with dysplasia, leukoplakia without dysplasia- NDB-UFES Dataset)Normal vs OSCC & OSCC vs leukoplakia with dysplasia vs leukoplakia without dysplasiaBest Performing Model (EfficientNetB3)Binary classification-Accuracy: 97.05Precision: 97.05%Recall (Sensitivity): 96.92%F1-score: 97.05%Multi-class classification-Accuracy: 97.16%Precision: 97.16%Recall: 97.16%F1-score: 97.16%Our Study9IndiaMobileNetV3, VGG16, Vision Transformer (ViT), ResNet-50, YOLOv11, Inception-V3 and EfficientNet-B0DLTotal images-2363 (train-1887, test-476)Normal vs Dysplasia vs OSCCBest Performing Model (EfficientNetB3)Multi-class classification-Accuracy: 97.48%Precision: 98%Recall: 97%F1-score: 97%Our Study34IndiaDL2080 (train-1661, test-419) digitized H&E-stained histopathological imagesMild dysplasia vs Moderate dysplasia vs Severe dysplasiaBest Performing Model (DenseNet-121)Multi-class classification-Accuracy: 95.47%Precision: 95%Recall: 95%F1-score: 95% Table 6Comparative analysis of previously published deep learning architectures for histopathological classification of normal epithelium, OED and OSCC.Table 6Authors, YearCountryAI architectureDL/ML/HybridImage countClassification approachResultsMaia et al. 202430BrazilCNNs (DenseNet-121), Transformer neural networks & Few-shot learning models (Siamese Network, Triplet Network, ProtoNet)DL3763 patches (P-NDB-UFES datasetNo Dysplasia vs Dysplasia vs OSCCBest Performing Model DenseNet-121Balanced Accuracy (BCC)- 91.91%Precision- 91.93%Recall- 91.93%Tafala et al. 202531MoroccoDeepPatchNet (DeepLabV3+ + ConvMixer) + Grad-CAMDL3763 (NDB-UFES dataset) 1020 [Oral Cavity-Derived Cancer (OCDC) dataset]No Dysplasia vs Dysplasia vs OSCCAccuracy- 86.71%Precision- 86.80%Recall- 86.71%F1-score- 86.75%Mandal et al. 202532IndiaResNet50 + Machine Learning classifiers (KNN, Linear SVM, Decision Tree, Random Forest, Logistic Regression, Naïve Bayes); Multimodal integration (histopathological image features + clinical-demographic data)Hybrid237 (NDB-UFES dataset)OSCC vs leukoplakia with dysplasia vs leukoplakia without dysplasiaKNN:Highest accuracy- 89%Linear SVM: Accuracy- 81%Akram et al. 202633PakistanDeep Visual Detection System (DVDS); EfficientNetB3; DenseNet121; ResNet50DL5192 (Normal vs OSCC- Kaggle)3763 (OSCC, leukoplakia with dysplasia, leukoplakia without dysplasia- NDB-UFES Dataset)Normal vs OSCC & OSCC vs leukoplakia with dysplasia vs leukoplakia without dysplasiaBest Performing Model (EfficientNetB3)Binary classification-Accuracy: 97.05Precision: 97.05%Recall (Sensitivity): 96.92%F1-score: 97.05%Multi-class classification-Accuracy: 97.16%Precision: 97.16%Recall: 97.16%F1-score: 97.16%Our Study9IndiaMobileNetV3, VGG16, Vision Transformer (ViT), ResNet-50, YOLOv11, Inception-V3 and EfficientNet-B0DLTotal images-2363 (train-1887, test-476)Normal vs Dysplasia vs OSCCBest Performing Model (EfficientNetB3)Multi-class classification-Accuracy: 97.48%Precision: 98%Recall: 97%F1-score: 97%Our Study34IndiaDL2080 (train-1661, test-419) digitized H&E-stained histopathological imagesMild dysplasia vs Moderate dysplasia vs Severe dysplasiaBest Performing Model (DenseNet-121)Multi-class classification-Accuracy: 95.47%Precision: 95%Recall: 95%F1-score: 95% Comparative analysis of previously published deep learning architectures for histopathological classification of normal epithelium, OED and OSCC. Collectively, these comparisons demonstrate that the findings of the present study are consistent with, and in some cases exceed, previously reported performances. The systematic benchmarking of multiple architectures further highlights the effectiveness of EfficientNet-based models in capturing complex histopathological features, supporting their potential utility for automated diagnostic support in digital oral pathology. Factors contributing to the superior performance of EfficientNet-B0 The superior performance of EfficientNet-B0 observed in the present study may be attributed to its compound scaling strategy, which simultaneously balances network depth, width and input resolution to achieve more efficient feature extraction and representation. This balanced scaling enables the model to capture both fine-grained cellular details and broader tissue-level patterns that are critical for distinguishing between normal epithelium, dysplasia and OSCC in histopathological images. Additionally, the parameter-efficient design of the architecture and its effective reuse of learned features enhance training stability and generalization capability, enabling the model to achieve high classification accuracy even when trained on relatively limited datasets. Previous work Previous work by the authors developed a YOLOv8-based deep learning model for the automated detection and classification of oral potentially malignant disorders and oral cancer using intraoral clinical images, demonstrating the potential of artificial intelligence driven approaches for non-invasive clinical screening and early identification of suspicious lesions.9 DenseNet-121 has also been reported to provide a reliable and generalizable framework for automated grading of oral epithelial dysplasia, contributing to improved diagnostic precision, reduced observer bias, and potential integration into digital pathology workflows for early detection and risk stratification of oral precancerous lesions.34 In clinical practice, AI-assisted image analysis could support the rapid evaluation of histopathological slides by enabling automated detection of regions that warrant closer examination. This capability may be especially valuable in high-volume diagnostic laboratories, where increasing workloads and limited time can influence diagnostic efficiency and turnaround time. Moreover, the integration of deep learning models into digital pathology platforms may facilitate a more objective assessment of both precancerous and cancerous lesions. Consequently, AI-assisted systems have the potential to support the expertise of oral pathologists by enabling earlier detection, improving diagnostic consistency and improving the overall efficiency of oral cancer screening and diagnostic workflows. Limitations and future directions Despite the promising findings observed in the present study, certain limitations should be acknowledged. The dataset utilized in this investigation was derived from a single institution, which may restrict the diversity of histopathological patterns, staining variations, scanner specifications and tissue processing protocols represented within the training and testing datasets. As histopathological image characteristics can vary considerably across institutions due to differences in slide preparation, staining intensity, image acquisition systems and laboratory workflows, the generalizability of the developed models to broader clinical environments may therefore be limited. In addition, the present study was conducted using patch-based histopathological image analysis rather than whole-slide image evaluation. Although patch-based approaches facilitate efficient model training and feature extraction, they may not fully capture the complete architectural context and spatial heterogeneity observed in routine pathological assessment. Furthermore, despite the implementation of augmentation strategies, weighted optimization and balanced sampling techniques, the relatively lower representation of carcinoma cases remains an inherent challenge frequently encountered in medical imaging datasets and may potentially influence model learning behavior and classification performance. Future investigations should therefore focus on the development and validation of large-scale multicenter datasets with standardized annotation protocols and external validation cohorts. Integration of whole-slide imaging, explainable AI techniques, calibration analysis and prospective clinical evaluation may further enhance interpretability, robustness and clinical applicability of AI-assisted oral histopathological diagnostic systems. Conclusion The present study demonstrated the effectiveness of deep learning based approaches for multiclass histopathological classification of normal oral epithelium, OED and OSCC. Among the evaluated architectures, EfficientNet-B0 achieved the highest overall performance, suggesting its enhanced capability to capture both cellular-level and tissue-level morphological characteristics relevant to oral epithelial lesion classification. The incorporation of stain normalization, augmentation strategies, weighted optimization and standardized benchmarking conditions further contributed to improved model robustness and comparative evaluation across architectures. Nevertheless, the findings of the present study are constrained by the use of a single-center, patch-based dataset and the absence of external multicenter validation, which may limit broader clinical generalizability. Accordingly, the proposed framework should be regarded as an adjunctive decision-support system designed to complement the expertise of pathologists and enhance diagnostic consistency, rather than function as an independent replacement for expert histopathological evaluation. Overall, the study provides a comprehensive benchmarking framework for future AI-assisted oral histopathological research and highlights the potential translational value of deep learning in automated screening and early detection of oral precancerous and malignant lesions. Future investigations involving larger multicenter datasets, whole-slide imaging, explainable AI integration and prospective clinical validation may further enhance the generalizability and clinical applicability of these approaches. Consent to participate Informed consent was obtained from all individual participants included in the study. CRediT authorship contribution statement Achla Bharti: Writing- original draft, Validation, Supervision, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Nikita Kashyap: Writing- original draft, Validation, Supervision, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Mala Kamboj: Writing- original draft, Validation, Supervision, Methodology, Investigation, Formal analysis, Data curation, Project administration, Software, Conceptualization. Kamaldeep Joshi: Writing-review & editing, Methodology, Investigation, Supervision, Software, Conceptualization. Sahil Hooda: Writing-review & editing, Supervision, Software, Conceptualization. Deepika Mishra: Writing-review & editing, Validation, Supervision, Methodology, Investigation, Formal analysis, Project administration, Conceptualization. Harpreet Singh: Writing-review & editing, Validation, Supervision, Methodology, Formal analysis. Debnath Pal: Writing-review & editing, Supervision, Software, Conceptualization. Ethics approval All procedures performed in this study involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. The study was covered by the Institutional Review Board. Consent for publication Consent for publication was obtained from all individual participants for whom identifying information is uniquely included in this manuscript. Human subjects statement The study utilized anonymized retrospective data and ethical approval was obtained from the Institutional Ethics Committee (Approval No- IRB No.: PGIDS/BHRC/23/15). Patient consent was waived due to the use of de-identified data. Funding The authors gratefully acknowledge the support of the Indian Council of Medical Research (ICMR) under the National Task Force project titled “Generate Medical Imaging Datasets (MIDAS) for tobacco induced oral lesions to enhance research in health in India” (Ref. No. BMI/014(03)/TF/2023, ID No. 2023-20012). Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.