Skip to main content
  • The introduction highlights the global impact of oral cancer and the alarming increase in mortality rates over time. While survival rates vary, early diagnosis is crucial for improving outcomes. The challenge of late-stage diagnoses, especially in remote areas, underscores the need for a simple, low-cost diagnostic tool. Oral squamous cell carcinoma (OSCC) is the predominant form of oral cancer, often originating from potentially malignant disorders (OPMDs). Detecting OPMDs early is essential, as some can progress to malignancy. Conventional visual assessment forms the basis of many screening programs, but it has limitations.
  • Artificial Intelligence (AI) has emerged as a potential solution for lesion detection. AI, modeled after human thought processes, shows promise despite its current limitations. Deep Convolutional Neural Network (DCNN) models have successfully detected precancerous skin lesions, achieving high accuracy. Some studies have used AI-based systems with pre-collected images, addressing limited access to care. The goal is to create an AI system that can be operated by frontline healthcare workers, including those without formal training, and can connect with remote specialists for informed diagnoses. This approach is expected to improve screening by distinguishing potentially malignant disorders from benign lesions, enabling early treatment.
  • The challenge lies in the computational demands of DCNN models, particularly for smartphone applications. Google’s efforts have led to the development of efficient models like MobileNet and EfficientNet, designed to balance accuracy and efficiency. Despite limited research in the field, this study aims to assess the accuracy of DCNN models, specifically EfficientNetV2 and MobileNetV3, in detecting and distinguishing precancerous oral lesions using pre-collected images. The hypothesis is that these models can accurately categorize oral lesions into different classes, aiding in early diagnosis and treatment.
  • The study found that EfficientNetV2 and MobileNetV3 achieved accurate identification (82-84%) of lesions in Classes 2 and 3, while accuracy was lower for Class 1 lesions (64-63%). This discrepancy suggests the models might excel at detecting severe lesions, highlighting a need to improve identification of milder or non-neoplastic lesions. Additional analyses using confusion matrices and ROC curves confirmed varying accuracy levels across lesion classes, with promising performance indicated by high AUC values of 0.88.
  • Prior research explored AI and computer vision for oral cancer diagnosis, demonstrating the feasibility of photographic imaging in assessing malignant disorders. Other studies revealed the potential of Convolutional Neural Networks (CNNs), achieving high AUC values. This proof-of-concept study demonstrated the feasibility of Deep Convolutional Neural Networks (MobileNetV3 and EfficientNetV2) for pre-collected oral lesion classification. While results are promising, further model development, especially for non-neoplastic lesions, is necessary.
  • The study’s outcomes could lead to a smartphone app for automated premalignant oral lesion diagnosis, utilizing portable image collection, computation, and data transmission. Challenges include accurate lesion identification within a focused field of view, and the need for further research to consolidate findings for clinical application.
  • The study acknowledges limitations, such as models struggling with certain lesion classes, not utilizing textural filters to assess baseline model performance, and lacking differentiation between lesion subtypes. Future work could address these limitations and enhance the understanding of model performance.
  • The study highlights the capacity of AI to improve remote oral lesion screening, aiding underserved populations. The proof-of-concept study effectively showcased the potential of AI, specifically MobileNetV3 and EfficientNetV2, to classify and recognize oral lesions. While the models showed promise in distinguishing between certain lesion types, their accuracy in detecting non-neoplastic lesions requires further improvement.


▪️- Further reading:
Malignant and non-malignant oral lesions classification and diagnosis with deep neural networks

Mahmoud H. Al-Johani

Author Mahmoud H. Al-Johani

More posts by Mahmoud H. Al-Johani

Leave a Reply