Two-step pipeline for oral diseases detection and classification: a deep learning approach

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Early detection of oral diseases, including oral potentially malignant disorders (OPMD) and oral squamous cell carcinoma (OSCC), is crucial for improving patient outcomes.
• Current diagnostic methods can be subjective and time-consuming, highlighting the need for objective and efficient tools.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) pipeline integrating object detection and classification models for the early identification and differentiation of oral diseases.
• To assess the performance of a two-step AI approach in analyzing clinical images of OPMD and OSCC.

Main Methods:

• A retrospective cross-sectional study using 773 baseline and 132 external validation oral clinical images.
• Development of YOLOv11 object detection models and MobileNetV2 classification models with various data augmentation and training strategies.
• Integration of the best performing detector and classifier into a two-step pipeline for disease identification.

Main Results:

• The best YOLOv11 model achieved a mean average precision (mAP50) of 0.820.
• The best MobileNetV2 classifier demonstrated an accuracy of 0.846 on the baseline dataset and 0.850 on external validation.
• The integrated two-step pipeline achieved an accuracy of 0.784 on the baseline test set and 0.863 on the external validation set, with high AUC-ROC values (0.811 and 0.934, respectively).

Conclusions:

• The developed AI pipeline shows feasibility for supporting early oral disease diagnosis.
• The AI models captured discriminative patterns for OPMD and OSCC, although some class overlap was observed.
• Caution is advised regarding the performance metrics of two-step approaches due to potential exclusion of images missed during the detection phase.