Deep OCT image compression with convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Optical coherence tomography (OCT) is crucial for diagnosing retinal diseases.
• High resolution OCT images require significant storage and bandwidth, hindering efficient data management.
• Existing compression methods may compromise diagnostic quality.

Purpose of the Study:

• To develop an end-to-end image compression framework for OCT using convolutional neural networks (CNNs).
• To achieve high compression ratios while preserving diagnostic image quality.
• To enable faster and more efficient storage and tele-transfer of retinal OCT images.

Main Methods:

• Developed a framework with three modules: data preprocessing, compression CNNs, and reconstruction CNNs.
• Incorporated skip connections with quantization for preserving fine-structure information.
• Utilized a dual-component objective function (adversarial discriminator and MS-SSIM penalty) for training.

Main Results:

• Achieved an image size compression ratio as high as 80.
• Reconstructed images maintained over 99% similarity (MS-SSIM) at 40x compression.
• 80x compressed images showed comparable quality to state-of-the-art methods at 20x compression.
• Demonstrated superior performance over other methods in MS-SSIM and visualization, especially at higher compression ratios.
• Compression and reconstruction were rapid, averaging 0.015 seconds per image.

Conclusions:

• The proposed CNN-based framework offers a promising solution for medical image compression.
• Deep neural networks can be effectively customized for OCT image compression.
• The framework significantly enhances efficient storage and tele-transfer of ophthalmic OCT data.