Learnable despeckling framework for optical coherence tomography images.

Area of Science:

• Biomedical Imaging
• Medical Image Analysis
• Computational Imaging

Background:

• Optical Coherence Tomography (OCT) provides high-resolution biomedical imaging but is degraded by speckle noise.
• Digital filters can improve OCT image quality, but selecting the optimal filter is challenging, especially for diverse dermatological tissues.
• Hardware modifications for image enhancement in clinical OCT devices are often prohibitively expensive.

Purpose of the Study:

• To propose an expandable and learnable framework, termed LDF, for automated speckle reduction in OCT images.
• To develop a system that learns a quantitative figure of merit (FOM) to assess filter effectiveness.
• To create a flexible framework capable of incorporating various despeckling algorithms.

Main Methods:

• Developed a learnable despeckling (LDF) framework comprising an autoencoder neural network and a filter classifier.
• The autoencoder learns a figure of merit (FOM) using image quality metrics like SNR, CNR, ENL, EPI, and MSSIM.
• The filter classifier evaluates sliding window, adaptive statistical, and edge-preserved correlation-based filters.

Main Results:

• The LDF framework learns a quantitative figure of merit (FOM) for effective speckle reduction.
• The system is designed to be learnable, improving performance with each image set.
• The expandable architecture allows for the integration of diverse despeckling algorithms.

Conclusions:

• LDF offers an automated and adaptable solution for speckle noise reduction in OCT images.
• The framework's learnable nature enhances its performance over time.
• LDF provides a cost-effective alternative to hardware upgrades for improving clinical OCT image quality.