[Scale-invariant feature-enhanced deep learning framework for oral mucosal lesion segmentation].

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Context:

• Accurate segmentation of oral mucosal lesions is crucial for diagnosis and treatment planning.
• Existing segmentation models often struggle with the subtle variations and boundaries of these lesions.
• Deep learning offers potential for improved automated analysis of medical images.

Purpose:

• To develop and evaluate PixelSIFT-UNet, a novel semantic segmentation model for enhanced oral mucosal lesion segmentation.
• To integrate the Scale-Invariant Feature Transform (SIFT) algorithm with a deep learning architecture (UNet).
• To compare the performance of PixelSIFT-UNet against traditional segmentation models like U-Net and PSPNet.

Summary:

• A dataset of 838 oral mucosal disease images was used to train and test the PixelSIFT-UNet model with VGG-16 and ResNet-50 backbones.
• The model achieved superior segmentation performance, with the ResNet-50 backbone yielding Dice coefficient of 0.668, mIoU of 0.733, mPA of 0.872, and Precision of 0.817.
• PixelSIFT-UNet demonstrated significant improvements over conventional U-Net and PSPNet models in segmenting oral lichen planus, leukoplakia, and submucous fibrosis.

Impact:

• PixelSIFT-UNet provides a more accurate and robust tool for the automated segmentation of oral mucosal lesions.
• The model's enhanced performance can aid clinicians in more precise diagnosis and treatment monitoring.
• This advancement contributes to the growing field of AI-driven medical image analysis for improved patient outcomes.