Uncertainty-Driven Parallel Transformer-Based Segmentation for Oral Disease Dataset.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Oral Pathology

Background:

• Accurate oral disease segmentation is hindered by disease variability, indistinct lesion boundaries, and limited public datasets.
• Existing methods struggle with the diverse visual characteristics and ambiguous edges of oral lesions.

Purpose of the Study:

• To develop an advanced deep learning model for accurate segmentation of multiple oral diseases.
• To introduce a novel network architecture and an uncertainty-driven loss function to enhance segmentation performance.
• To create and release a large-scale dataset for oral disease segmentation research.

Main Methods:

• Developed Oralformer, a U-shaped encoder-decoder network utilizing a parallel LC-block combining local-window self-attention (LWSA) and channel-wise convolution (CWC).
• Implemented an uncertainty-driven self-adaptive loss function to improve focus on ambiguous lesion boundaries.
• Constructed a large-scale oral disease segmentation (ODS) dataset with 2602 image pairs covering plaque, calculus, and caries.

Main Results:

• Oralformer achieved state-of-the-art segmentation accuracy across six challenging datasets.
• Demonstrated superior generalizability and real-time segmentation efficiency at 35 frames per second.
• The uncertainty-driven loss function effectively improved segmentation of difficult-to-identify lesion edges.

Conclusions:

• Oralformer presents a significant advancement in automated oral disease segmentation.
• The developed methods and dataset are expected to accelerate research and clinical applications in oral health.
• Publicly available code and dataset facilitate further development and validation.