LS-Net: lightweight segmentation network for dermatological epidermal segmentation in optical coherence tomography...

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) offers non-invasive skin evaluation for diagnosing diseases.
• Accurate epidermal segmentation in OCT is challenging due to skin's diverse morphology, limiting its clinical utility.
• Existing segmentation methods struggle with speed and accuracy.

Purpose of the Study:

• To develop a novel, efficient deep learning model for precise epidermal segmentation in OCT images.
• To address the limitations of current methods in handling skin morphological diversity.
• To enhance the diagnostic capabilities of OCT in dermatology.

Main Methods:

• Introduced LS-Net, a lightweight segmentation network combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).
• Incorporated a depth-wise convolutional transformer for spatial contextualization and a squeeze-and-excitation block for feature recalibration.
• Validated the model on a custom dataset across various skin sites (face, palm) under clinical conditions.

Main Results:

• LS-Net achieved high segmentation accuracy (mean Dice: 0.9624, mean IoU: 0.9468).
• The network demonstrated significantly reduced computational demands (1.131 G floating point operations).
• Outperformed existing segmentation methods in accuracy and efficiency.

Conclusions:

• LS-Net provides precise and computationally efficient epidermal segmentation for OCT.
• The model's effectiveness across diverse skin sites and conditions supports its clinical applicability.
• LS-Net enhances OCT's diagnostic value in dermatological practice.