PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
May 17, 2026Dental Research Journal0 citationsOpen Access

Evaluating the diagnostic accuracy of neural network models in detecting oral potentially malignant disorders and oral cancer using mobile photographs: An umbrella review

View Full Paper
PKP Madan KumarRSRajeshwari SelvamSSSasidharan Sivakumar

Key Points

  • The aim is to evaluate the comparative performance of various AI models in detecting oral cancer and oral potentially malignant disorders.
  • An umbrella review was conducted using six systematic reviews from databases including Medline, Web of Science, Scopus, and EMBASE.
  • The risk of bias was assessed using the Joanna Briggs Institute’s assessment tool.
  • Pooled sensitivity for AI-based detection ranged from 88% to 92%, with diagnostic odds ratios between 114 and 2549.
  • Deep learning models such as EfficientNet and ResNet showed high diagnostic accuracy, while hybrid models like MLSO + SVM demonstrated promise.
  • Substantial heterogeneity was noted across studies, often with I² greater than 85%.

Abstract

ABSTRACT Background: Oral cancer (OC) and oral potentially malignant disorders (OPMDs) remain major global public health challenges, particularly in low-and middle-income countries. Although early detection substantially improves prognosis, limited healthcare infrastructure restricts timely diagnosis. Artificial intelligence (AI) enabled, mobile phone-based diagnostic systems offer a promising, accessible solution, and multiple systematic reviews have demonstrated their potential. However, uncertainty persists regarding the comparative performance of AI models across diverse real-world settings. Aim: An umbrella review was aimed at evaluating the comparative performance of different AI models in detecting OC and OPMD. Materials and Methods: This research identified six systematic reviews from databases such as Medline (via PubMed), Web of Science, Scopus, and EMBASE through October 2024 which were checked at the title, abstract, and full-text levels. The risk of bias (ROB) was then assessed using the Joanna Briggs Institute’s ROB assessment tool. Results: Across included reviews, pooled sensitivity and specificity for AI-based detection ranged from 88% to 92%, with reported diagnostic odds ratios ranging from 114 to 2549, indicating strong discriminatory performance. Deep learning architectures such as EfficientNet and ResNet consistently demonstrated high diagnostic accuracy, while hybrid approaches (e.g., MLSO + SVM) showed promising performance in selected analyses. However, substantial heterogeneity was observed across studies ( I 2 often >85%), reflecting variability in populations, image acquisition protocols, and model architectures. Conclusion: Deep learning models like EfficientNet and ResNet are favored in clinical diagnostics for their exceptional performance and adaptability. Hybrid approaches, such as MLSO + SVM, also show great potential by combining the strengths of traditional and modern methods effectively.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Kumar et al. (2026) studied this question.

synapsesocial.com/papers/6a095b3f7880e6d24efe10b5https://doi.org/10.4103/drj.drj_172_25
Ask AI
Helpful
Bookmark
Share
View Full Paper