Oral squamous cell carcinoma (OSCC) is the most common malignant oral cavity tumor and one of the most important public health problems in the world.[1] According to recent estimates worldwide, cancers of the lip and oral cavity are responsible for hundreds of thousands of new cancer cases and deaths each year, especially in South and South-East Asia.[2] In 2022, approximately 177,258 new oral cancer cases and 98,735 deaths were reported in South and South-East Asia alone.[2] Pakistan is among the countries in this region with a comparatively high burden, with an age-standardized incidence rate of 12.07 per 100,000 males and mortality rate of 7.74 per 100,000 males..[2, 3]  This worldwide burden has also risen over the last 30 years: analysis of the Global Burden of Disease data showed that in 1990, the age-standardized incidence rate of oral cancer was 3.26 cases per 100,000, rising to 5.34 cases per 100,000 in 2021.[4]

Histopathological examination continues to be the mainstay for the diagnosis and biological features of OSCC.[5] In addition to the diagnosis of malignancy, histopathological grading gives information about the degree of tumor differentiation and forms part of the evaluation of the behaviour and prognosis of the tumor.[6] However, conventional grading involves a microscopic evaluation of several cell and architectural features, which can be subjective, especially when the tumor has morphologically heterogeneous or borderline features.[7] Traditionally, the WHO histopathological approach has focused on the extent of tumor differentiation, and more recent studies have put emphasis on other morphological parameters which appear to be of prognostic significance, such as stromal reaction, tumor-cell dissociation, and growth pattern.[8, 9] Therefore, consistent and reproducible grading is an important challenge in routine oral pathology practice.

The recent emergence of digital pathology and Artificial Intelligence (AI) has opened new avenues for objective assessment of histopathology images.[10] Machine-learning and deep-learning algorithms can be used to analyse digitized tissue sections at a high magnification, detect cellular or architectural patterns, and offer reproducible quantitative assessments.[11] There is considerable evidence gained in the last few years that indicates a lot of diagnostic potential. A systematic review of AI applications on histopathological images reported a range of accuracies from 89.47% to 100%, and a systematic review and meta-analysis of deep learning approaches for the histopathology of OSCC reported a pooled sensitivity of 98% and specificity of 93% for OSCC detection.[12, 13] Notably, AI has gone beyond merely detecting OSCC; an AI-based system using EfficientNet for OSCC histological grading showed an accuracy of 98.1% in a large reference set of images and 91.4% accuracy in an independent image set, indicating the possibility of computational methods to aid the assessment of tumor grade.[14]

However, while these studies have yielded encouraging results, the majority have investigated AI as a standalone diagnostic tool or primarily as a way to identify only between malignant and non-malignant. There have been fewer studies that directly compared AI-assisted grading with conventional microscopic grading of the same OSCC specimens and a defined histopathological reference standard. This distinction is clinically relevant because there is not a strong correlation between the accuracy of diagnosis of OSCC and the accuracy of tumor grading. Further, recent reviews persist in highlighting issues such as small sample sizes, methodological variation, lack of external validation, and possible model bias.

A direct comparison with standard microscopy could then help to determine if AI can offer the same, complementary, or conflicting information for grading and if it can enhance the consistency and repeatability of standard oral cancer assessment. Hence, the present study was planned to fill this clinically relevant gap by comparing the diagnostic accuracy of AI-assisted histopathological analysis with conventional light microscopy for grading OSCC. The aim of the study was to define and compare the sensitivity, specificity, predictive values, overall diagnostic accuracy, and agreement of AI-assisted histopathological grading with conventional histopathological grading, with reference to the expert histopathological assessment.

The study was a comparative diagnostic accuracy study carried out in the Department of Oral Pathology, Liaquat University of Medical and Health Sciences (LUMHS), Jamshoro, Sindh, Pakistan. The study was carried out during six months, from 1st January, 2026 to 30th June, 2026. Sample size was determined using OpenEpi version 3.01 for estimation of a single proportion. The accuracy used for the calculation was the diagnostic accuracy of the diagnostic performance of an AI system for histopathological evaluation of OSCC reported by Xu et al., who used an Efficient Net-based AI system and reported 91.4% in an independent image set. The minimum sample size required was about 121 cases, with an expected accuracy of 91.4%, a 95% confidence level, and an absolute precision of 5%.[14] Potential non-evaluable or technically poor slides caused the sample size to be increased by about 10% to a target sample of 134 OSCC cases. A consecutive sampling technique was used. Patients of either gender, all age groups (adults), and those with a histopathologically established diagnosis of primary oral squamous cell carcinoma were included in the study. Cases were eligible if adequate formalin-fixed, paraffin-embedded tissue blocks and hematoxylin and eosin (H&E) stained sections were available for evaluation. The selection of specimens included only those with adequate viable tumor material for evaluation of histological differentiation and grading with conventional microscopy and AI. Samples with inadequate, poorly preserved, extensively necrotic, autolyzed, and severely crushed tissue that made reliable histopathological diagnosis impossible were excluded. Recurrent tumors, non-diagnostic slides, due to severe staining or processing artifacts, and insufficient tumor tissue were also not included. If there was no tissue block available or no good digitized image, then these cases were not included in the final analysis. Eligible cases of OSCC were identified from the histopathology records during the study period after ethical approval from the respective institutional research ethics committee. A structured data collection proforma was used to record demographic and clinicopathological information such as age, sex, anatomical site, and histological grade. Conventional light microscopy (H&E staining) was used by experienced oral pathologists to review the original H&E-stained sections. The tumours were classified into well-differentiated, moderately differentiated, or poorly differentiated OSCC, and histological grading was done as per the predefined grading criteria. Representative H&E-stained sections from the same cases were digitized at appropriate resolution by a whole slide imaging system for AI-assisted assessment. A selected AI-based histopathological image-analysis system was used to process the digitized images. The images were preprocessed and quality-controlled prior to the analysis, while the AI system categorized the tumor into the given histological grades. The AI assessment was carried out independently of the conventional microscopic assessment to reduce incorporation and interpretation bias. Two experienced oral pathologists, who had reviewed the pertinent H&E sections independently, agreed on the final diagnosis in cases of disagreement. The reference diagnosis remained blinded to the AI-generated diagnosis in the initial evaluation. The AI-generated grade and traditional microscopic grade were then compared to the reference standard. This method was chosen as it has been shown that AI systems can be effective in OSCC histopathological analysis; for instance, an Efficient Net-based model has yielded an accuracy of 98.1% on the TCGA dataset and 91.4% on an independent dataset.[14] All data collected were entered and analyzed with SPSS version 26. Demographic and clinicopathological features of the study group were summarized using descriptive statistics. The continuous variables were reported as mean±SD, while categorical variables including sex, tumor site, histological grade, and AI classification were reported as frequencies and percentages. The Shapiro - Wilk test was used to assess normality of continuous variables. The ability of the AI-assisted grading and traditional microscopy methods to diagnose the pathological state was evaluated against the consensus histopathological reference standard. The PPV, NPV, overall D, and 95% CIs were obtained for the identification of the relevant tumor-grade categories. Receiver operating characteristic (ROC) curves were plotted, and the area under the ROC curve (AUC) was computed to measure discriminatory performance. The degree of agreement between the AI-assisted grading results, the conventional microscopy results, and the reference result was evaluated by Cohen's kappa coefficient in categorical classifications. Weighted kappa was used for comparisons with more than two grading categories. The differences in diagnostic approaches were evaluated by suitable paired categorical tests like McNemar's test for paired binary data. A p-value <0.05 was considered statistically significant.

A total of 134 patients with histopathologically confirmed OSCC were included in the study. The average age of the participants was 57.8 ± 12.4 years. The patients were mostly aged 41-60 years, and males were predominant in the study population. The most common site of involvement was the tongue, followed by the buccal mucosa. The reference histopathological assessment indicated that most of the tumors were well differentiated, while moderately and poorly differentiated tumors were the next most common. (Table 1)

Overall, the distribution of tumor grades when evaluated by conventional microscopy and AI-assisted histopathology was similar to the reference standard. No significant differences in the number of cases in each grade were seen between the two methods, and the most uniform were the poorly differentiated cases. (Table 2)

Both diagnostic methods had good diagnostic performance when compared to the reference standard. AI-assisted histopathology had a slightly higher sensitivity, specificity, PPV, NPV, and overall accuracy when compared to conventional microscopy. The confidence intervals for accuracy further showed high accuracy of both methods. (Table 3)

The almost perfect agreement between the various diagnostic methods and the reference standard was shown by agreement analysis. AI-assisted grading demonstrated slightly greater agreement, and there was almost perfect agreement between AI-assisted grading and conventional microscopy. The results of the three comparisons were statistically significant. (Table 4)

ROC analysis demonstrated excellent discriminatory performance for both methods. The AUC was slightly higher for AI-assisted histopathology compared to conventional microscopy, but the difference was not significant between the two AUCs, meaning the overall discriminatory performance was similar. (Table 5)

The comparison results by direct paired comparison demonstrated that the histological grade given by AI-assisted histopathology and conventional microscopy was identical in the vast majority of cases. In the remaining discordant cases, the discrepancies were only one histological grade, and there was no two-grade discrepancy. The paired comparison was not significant, but the weighted kappa showed an extremely good agreement between the two methods. (Table 6)

Table 1. Demographic and clinicopathological characteristics of the study participants (n=134)

Table 2. Comparison of conventional microscopy and AI-assisted grading with the reference standard

Table 3. Diagnostic performance of conventional microscopy and AI-assisted histopathology

Table 4. Agreement between diagnostic methods and the reference standard

Table 5. ROC analysis of AI-assisted and conventional grading

Table 6. Paired comparison of grading classifications between AI-assisted histopathology and

conventional microscopy

.

In the present study, we have shown that artificial intelligence (AI)-based histopathological grading of oral squamous cell carcinoma (OSCC) shows high diagnostic performance and good correlation with the expert histopathological grading. AI-powered grading systems achieved a sensitivity of 95.1%, specificity of 96.1%, positive predictive value of 94.9%, negative predictive value of 96.3%, and overall accuracy of 95.5% in 134 OSCC cases. The conventional microscopy also showed high performance, achieving an overall accuracy of 94.0%. Nearly complete agreement with the reference standard (κ=0.92) was obtained with AI, compared with κ=0.90 with conventional microscopy, and the slightly better performance of AI was matched by the quicker time to analysis. The results indicate that AI-based analysis can perform grading as well as experts, with high consistency and with performance similar to the standard microscopic grading. The results are similar to those obtained by Xu et al. (2022), who proposed an auxiliary diagnostic system for the histopathological evaluation of OSCC using an Efficient Net network. In the TCGA dataset, their model performed with an accuracy of 98.1% and an AUROC of 0.998; in an independent dataset they maintained 91.4% accuracy and an AUROC of 0.992. Overall, the accuracy of AI in the present study was 95.5%, and the AUC was 0.963, which is not far from the level of performance of AI in the other study; thus, they are clinically relevant, as both studies showed that AI has high-level performance in the assessment of histopathological OSCC.[14] Likewise, Yang et al. (2022) trained a deep-learning model using histopathological images, achieving a sensitivity of 98%, specificity of 92%, positive predictive value of 92.4%, negative predictive value of 97.8%, and F1 score of 95.1%. Critically, their experiment involving reader assistance showed that AI can enhance the performance of both junior and senior pathologists and that it can also speed up the image assessment of junior pathologists. These observations are corroborated by the present findings, which demonstrated that AI-based grading had a slightly higher accuracy in the diagnosis than conventional microscopy and showed almost perfect correlation with the reference diagnosis. The present study, however, explicitly analyzed tumor-grade classification instead of only distinguishing OSCC from non-OSCC tissue, as in Yang et al.[15] Similar results were obtained by Panigrahi et al. (2023), who studied deep transfer learning for classification of histopathological images of OSCC. Their work showed that transfer-learning approaches are feasible for automated histopathological classification, in support of the idea that morphological features discernible on H&E sections can be captured computationally. The level of agreement seen in the current research further accentuates the possibilities of AI to deliver repeatable pattern interpretation of morphology. However, comparisons should be taken with a grain of salt since the datasets, image preparation, and model architecture and classification endpoints vary between the two studies.[16] Sukegawa et al. (2023) tested deep-learning classifiers for OSCC histopathological images and examined whether artificial intelligence can help the oral pathologist. The highest-performing VGG16 model achieved an accuracy of 86.22% and an AUC of 0.9602 for the diagnostic performance of the oral pathologists, who also received supplementary information that greatly enhanced this diagnostic performance. The overall accuracy of the present AI-assisted approach is 95.5%, and the AUC is 0.963, whereas the accuracy and AUC of the previous method are 87.6% and 0.925, respectively. Both studies had broadly similar AUC values, suggesting good discriminatory capacity, but with a higher accuracy in the present study, which may have been due in part to the differences between the two studies in case selection, task definition, and reference-standard methodology. The results of both studies indicate that AI should be used as an add-on rather than a replacement for the clinical practice of oral pathologies.[17] A 2024 study by Albalawi and colleagues using an EfficientNet-based approach for histopathological OSCC detection also demonstrated the ability of deep-learning models to distinguish normal oral epithelium from OSCC on digitized histological images. They used 1,224 images of 230 patients in their dataset, an EfficientNetB3 model and image augmentation and optimization strategies. While their endpoint was different from the one in the present study, their results support the notion that AI can be used to recognize diagnostically relevant morphological patterns in H&E-stained tissue. The current research was extended from disease detection to the more challenging task of distinguishing OSCC grades.[18] The results of the present study are also confirmed by the systematic review and meta-analysis conducted by Pirayesh et al. in 2024 that aimed to assess deep-learning-based image classification and segmentation of digital histopathology for OSCC. In 17 eligible studies, reported classification performance was very high, with pooled sensitivity of 98% and specificity of 93%, but only three studies had a low risk of bias for all of the QUADAS-2 domains. The sensitivity and specificity of the present study are thus in the range of those reported in the literature. Moreover, the current study provides evidence based on a direct comparison between AI and conventional microscopy, whereas the previous investigations focused solely on the evaluation of an AI model.[13] These results should also be taken into account together with the results of Khanagar et al. (2023), which synthesized the existing evidence on the use of AI in diagnosis, classification, and prediction of oral cancer in histopathological images. Their review revealed significant promise for AI in various histopathological applications, yet highlighted inconsistencies in data sets, algorithms, diagnostic goals, and validation methodologies. In the present study, these methodological issues were addressed by comparing both AI-assisted and conventional microscopic evaluation of the same cases of OSCC using the consensus diagnosis from experienced oral pathologists as the reference standard. This design facilitates a more direct interpretation of the high agreement of AI with the reference diagnosis within the context of the clinical diagnostic process.[19] Whole slide imaging (WSI) has been used more recently to develop a clustering-constrained attention multiple-instance learning (CLAM) model specifically for predicting the pathological grade of OSCC in a study by Shao et al., 2025. In the internal validation, their model obtained an average AUC of 0.86 across different pathological grades, and an AUC of 0.71 in external validation with CPTAC data. This study showed that AI-based classification had a significantly higher AUC of 0.963. It should be noted, however, that the difference should not be interpreted as proof of the superiority of one approach, since the models that were evaluated by Shao et al. were trained and tested on public databases, while the present investigation compared AI-assisted assessment with traditional microscopy in a specific clinical cohort. Despite their performance during external validation, however, it is a key point to remember that good performance in one set does not necessarily equal good performance in another institution, population, staining protocol, and scanner.[20] Ramya and Minu proposed a deep transformer encoder-based architecture that is specially designed for grading OSCC in 2025. They proposed the DeTr-DiGAtt framework that obtained an accuracy of 98.59%, a Dice score of 97.97%, and an intersection over union of 98.08%. The findings show the evolving complexity of AI models tailored for the 3-level grading of OSCC. The AI accuracy achieved in the present study (95.5%) was less than that previously reported for this model, but the type of evidence generated by the current design is a clinical case series in which the AI model's accuracy is compared to an expert reference diagnosis, since this model is tested directly against conventional microscopy. Overall, the two studies suggest that AI grading is technically achievable, but stress the need for validation in clinically realistic settings.[21] The present findings also have implications for the reproducibility of OSCC grading. The conventional histopathological grading relies on the interpretations of cellular differentiation and architectural features, and some tumors may give rise to inter-observer disagreement. AI has the potential to provide a consistent computational assessment once appropriately trained and validated. This study therefore has promising kappa values, especially as agreement between AI and the expert reference standard was slightly better than that between conventional microscopy and the expert reference standard. However, before routine use, the reproducibility of the method should be verified in various institutions, scanners, staining procedures, and patient groups. This is a major advantage of the present study because the same OSCC specimens have been used for direct comparison of the AI-assisted grading with conventional microscopy, with consensus expert grading as the reference standard. A significant number of earlier studies focused on diagnosis of OSCC in benign or non-cancerous tissue, while the more difficult setting of assigning the histological grade has been comparatively understudied. The present study thus provides evidence specifically in relation to the grade classification. In addition, sensitivity, specificity, predictive values, accuracy, ROC analysis, kappa statistics, and paired comparison offer a multidimensional evaluation of the performance of the diagnosis, not only by means of accuracy. Several limitations should nevertheless be considered. First, the sample of 134 cases was from a single institution, which may restrict the generalizability of the results to other samples and laboratory settings. Second, the study was limited to second opinions on already confirmed OSCC cases and thus did not test whether the AI could be used to differentiate between OSCC and benign lesions, potentially malignant disorders, or inflammatory conditions; this is a limitation of the study and does not imply that the AI will be capable of such discrimination in routine clinical practice. Third, the AI system may respond differently to different slides, depending on their staining qualities, their quality of scanning, their artifacts, and tumor variability. Fourth, while two experienced pathologists can give a de facto consensus standard, the grading itself is subject to some subjectivity. Lastly, there was no external multicenter validation.

AI-assisted histopathological grading demonstrated excellent diagnostic performance and almost perfect agreement with expert histopathological assessment in OSCC. However, there were no significant differences in the sensitivity, specificity, positive and negative predictive values, accuracy, or AUC found between AI and traditional microscopy. Importantly, over 93% of tumors were assigned the same grade using both methods, and there were no cases where there was a two-grade difference. These results validate the use of AI as a reliable, repeatable tool alongside traditional microscopy, which can help to increase the accuracy of diagnosis of OSCC and precision in their grading, as well as help pathologists with their daily work. Widespread clinical use will need to be supported by larger and more varied multicenter studies, external validation and evaluation of diagnostically challenging cases.