A systematic literature review of convolutional neural network and transfer learning approaches for smoker’s melanosis classification on intraoral images: methods, evidence quality, and readiness for clinical decision support

Background Smoker’s melanosis is a benign melanin hyperpigmentation of the oral mucosa, common in smokers and clinically useful both as a marker of tobacco exposure and as a lesion requiring differential diagnosis. Its severity is still graded by subjective visual indices, motivating objective, image-based approaches using deep learning. This review maps how convolutional neural networks, transfer learning, and the EfficientNet family have been applied to smoker’s melanosis and related oral mucosal lesions on intraoral images, appraises the quality of that evidence, and assesses its readiness for clinical decision support. Methods The review followed the PRISMA 2020 statement and the Kitchenham and Charters guidelines, addressing five research questions on architectures, transfer learning strategy, class imbalance and acquisition variability, evaluation protocols, and interpretability and clinical integration. Six databases (IEEE Xplore, ScienceDirect, PubMed/Medline, SpringerLink, Scopus, Google Scholar) were searched for 2019 to 2025 with additional citation searching; the last search was run on 28 March 2026. Risk of bias was appraised with PROBAST and reporting completeness against the CLAIM checklist; certainty per research question was rated using a GRADE-CERQual style scheme with pre-specified downgrading rules. The protocol was not registered in PROSPERO or any other prospective registry; eligibility criteria were fixed in writing before screening. Results From 440 identified records, 21 studies were included. Seven address the target condition directly, but only two develop a computational model; the remaining five are clinical observational studies. Transfer learning was the dominant strategy. Four of the five appraised primary studies were at high risk of bias in the analysis domain, reflecting small samples, reliance on overall accuracy, and absent calibration. Subject-wise, patient-level data splitting was applied in only one study, and none reported external cross-device validation. No published EfficientNet study has used intraoral images of smoker’s melanosis; the sole EfficientNet study identified used non-intraoral photographs of smoking scenes. Conclusions EfficientNet-based transfer learning is a promising but unproven direction for smoker’s melanosis. Progress requires a graded intraoral dataset with standardised acquisition, imbalance handling as a design constraint, clinically oriented evaluation with subject-wise splitting and calibration, and interpretability. We also set out image acquisition standards clinicians can adopt now.