Transfer Learning-Based Approach for Thickness Estimation on Optical Coherence Tomography of Varicose Veins

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Soft Tissue Mechanics

Background:

• Accurate mechanical characterization of veins is crucial for developing venous substitutes and understanding venous diseases.
• Key physical parameters like shape and thickness are challenging to determine in soft tissues.

Purpose of the Study:

• To propose and validate the TREE (Transfer learning-based approach for thickness Estimation) method for predicting vein segmentation maps and thickness.
• To leverage transfer learning to enhance the accuracy of vein thickness estimation.

Main Methods:

• The TREE method utilizes a one-encoder, two-decoder architecture trained sequentially using transfer learning.
• An encoder-decoder pair first predicts segmentation maps, followed by a second decoder trained for thickness prediction using a pre-trained encoder.
• A sensitive pattern detector (SPD) module was incorporated to extract semantic details and improve performance.
• Swept-source optical coherence tomography (SS-OCT) was used for imaging saphenous varicose veins.

Main Results:

• The TREE model achieved a segmentation accuracy of 0.993.
• The mean squared error for thickness estimation was 2.409 pixels.
• Performance metrics significantly outperformed existing state-of-the-art methods.

Conclusions:

• The TREE method demonstrates high accuracy in predicting vein segmentation and thickness from SS-OCT images.
• This approach offers a promising tool for in-depth mechanical characterization of veins.
• The findings contribute to advancements in venous disease research and the development of novel venous substitutes.