A Deep Learning AI Model for Histopathological Diagnosis and Grading of Mucoepidermoid Carcinoma of Salivary Glands
Clinical Snapshot
PICO Framework
| P — Population | Whole slide images (WSIs) of salivary gland tissue — 92 mucoepidermoid carcinoma (MEC) cases and 102 non-MEC cases — derived from archival pathology material at Cairo University Faculty of Dentistry |
| I — Intervention | A multiple instance learning (MIL)-based deep learning model performing (1) binary slide-level classification (MEC vs. non-MEC) and (2) three-class grading (low, intermediate, high grade MEC) |
| C — Comparator | Established histopathological diagnosis and grading by expert oral and maxillofacial pathologists serving as the reference standard |
| O — Outcomes | Classification accuracy, macro/weighted F1 score, macro precision, macro recall for the binary classification module; accuracy, weighted F1 score, weighted precision, weighted recall for the grading module |
Bottom Line
This proof-of-concept study presents a MIL-based deep learning model for histopathological diagnosis and grading of mucoepidermoid carcinoma (MEC) of salivary glands, trained and internally validated on 194 whole slide images from a single Egyptian institution. The binary classification module (MEC vs. non-MEC) achieved 90.62% accuracy, which is encouraging. However, the grading module — clinically the more consequential task — achieved only 73.3% accuracy without confidence intervals, making its real-world utility uncertain. Critical methodological gaps undermine confidence in these results: no external validation, no confidence intervals, no sensitivity/specificity reporting, undefined comparator group composition, and an underpowered test set. The reference standard methodology and blinding procedures are inadequately described, raising concerns about incorporation and review bias. MEC grade directly drives surgical and adjuvant treatment decisions; a model with unquantified precision at 73.3% accuracy cannot be recommended for independent clinical use. The authors appropriately conclude that AI should augment rather than replace pathologist judgement. This study represents an early-stage feasibility demonstration requiring substantial further development, prospective multicentre validation, and regulatory evaluation before any clinical translation. Senior clinicians and pathology services should not modify current diagnostic workflows based on these findings.
Key Findings
P Value: Not reported
Effect Size: Binary classification accuracy: 90.62%; macro F1 score: 90.62%; macro precision: 91.67%; macro recall: 91.18%. Grading model accuracy: 73.3%; weighted F1 score: 73.9%; weighted precision: 76.1%; weighted recall: 73.3%
Primary Outcome: Binary classification of WSIs into MEC vs. non-MEC, and three-class grading of MEC into low, intermediate, and high grade
Nnt Or Sensitivity: Macro recall (approximating mean sensitivity across classes) for binary classification: 91.18%. Class-specific sensitivity and specificity not reported. Grading model per-class sensitivity not reported. Positive and negative likelihood ratios not calculable from reported data.
Confidence Interval: Not reported for any metric
Clinical Application
Requires whole slide imaging infrastructure and digital pathology workflow integration. MIL-based models are computationally feasible on modern GPU hardware. However, the model is not publicly available, not externally validated, and not ready for clinical deployment. Significant further development, prospective validation, and regulatory approval would be required before clinical use. In Australia, digital pathology adoption is increasing but remains variable across public and private laboratory networks. The TGA would classify a diagnostic AI model of this nature as a Class IIb or Class III medical device under the Software as a Medical Device (SaMD) framework, requiring conformity assessment before clinical use. The Royal College of Pathologists of Australasia (RCPA) has published guidance on digital pathology and AI, emphasising that AI tools must be validated in the local population before adoption. MEC is managed in Australian tertiary head and neck oncology centres; grading directly influences decisions regarding adjuvant radiotherapy and surgical margins per current NHMRC-aligned multidisciplinary team guidelines. A grading accuracy of 73.3% without confidence intervals is insufficient to meet RCPA or TGA standards for clinical deployment. PBS reimbursement considerations are not directly applicable to a diagnostic AI tool of this nature. Potentially applicable to pathologists evaluating salivary gland biopsy or resection specimens where MEC is in the differential diagnosis. Most relevant in high-volume centres with digital pathology infrastructure. Not yet validated for independent clinical use.
Abstract
PURPOSE: This study aims to evaluate the effectiveness of deep learning AI model in the diagnosis and grading of mucoepidermoid carcinoma. METHODS: We developed a multiple instance learning based-deep learning model which consisted of two modules. One to perform slide-level classification of whole slide images into mucoepidermoid carcinoma and non-mucoepidermoid carcinoma classes and the second one to perform grading into low, intermediate, and high grades. A total of 194 whole slide images were included in the study. Of these, 92 corresponded to mucoepidermoid carcinoma, while the remaining 102 represented non-mucoepidermoid carcinoma cases. RESULTS: Our results showed that the classification model achieved 90.62% accuracy, 90.62% macro F1 score, 90.60% weighted F1 score, 91.67% macro precision, 91.18% macro recall. The grading model achieved an accuracy of 73.3%, weighted F1-score of 73.9%, weighted precision of 76.1% and weighted recall of 73.3%. CONCLUSION: The proposed deep learning classification model achieved promising performance in differentiating mucoepidermoid carcinoma cases. However, the grading model achieved an overall accuracy of 73.3%. Thus, depending on AI solely is still questionable, however it can be used to augment the work of the pathologist.
References
- 1.Abd El Ati, D. I., Fathy, S., Hamza, H. E., & Hamza, S. (2026). A deep learning AI model for histopathological diagnosis and grading of mucoepidermoid carcinoma of salivary glands. Head and Neck Pathology. https://doi.org/10.47852/bonviewAIA52026661
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service