An end-to-end implicit neural representation architecture for medical volume data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• Medical volume data is rapidly increasing, posing challenges for organization, storage, transmission, and rendering.
• Current methods struggle to efficiently manage petabyte-scale medical datasets.

Purpose of the Study:

• To develop an end-to-end deep learning architecture for efficient medical data compression.
• To balance high compression rates with superior reconstruction quality for medical imaging data.

Main Methods:

• An architecture combining downsampling, implicit neural representation (INR), and super-resolution (SR) modules was proposed.
• A trade-off point method was used to optimize module performance for compression and reconstruction.
• The method was experimentally validated on multi-parametric MRI data.

Main Results:

• Achieved a high compression rate of up to 97.5%.
• Maintained superior reconstruction accuracy with a Peak Signal-to-Noise Ratio (PSNR) of 40.05 dB and Structural Similarity Index (SSIM) of 0.96.
• Significantly reduced GPU memory requirements and processing time.

Conclusions:

• The proposed deep learning architecture offers a practical solution for handling large medical datasets.
• This approach effectively addresses the challenges of medical data volume growth.
• The method enables efficient storage, transmission, and manipulation of medical imaging data.