A Novel Implicit Neural Representation for Volume Data

TL;DR

This work tackles the practical problem of efficiently compressing high-resolution volumetric medical images using implicit neural representations. It introduces a three-module pipeline that uses Lanczos downsampling to LR, a SIREN-based implicit representation to encode the LR data, and SRDenseNet to reconstruct the HR volume, enabling high compression with reduced training time and memory on standard hardware. Empirical results on Visible Human CT data show higher reconstruction quality (PSNR) and faster training with lower GPU memory compared to baseline INR approaches, including SIREN alone, and outperform several existing volume-compression methods. The approach offers a flexible, configurable framework for efficient volume rendering and could facilitate faster clinical workflows and broader accessibility of high-resolution medical imaging data.

Abstract

The storage of medical images is one of the challenges in the medical imaging field. There are variable works that use implicit neural representation (INR) to compress volumetric medical images. However, there is room to improve the compression rate for volumetric medical images. Most of the INR techniques need a huge amount of GPU memory and a long training time for high-quality medical volume rendering. In this paper, we present a novel implicit neural representation to compress volume data using our proposed architecture, that is, the Lanczos downsampling scheme, SIREN deep network, and SRDenseNet high-resolution scheme. Our architecture can effectively reduce training time, and gain a high compression rate while retaining the final rendering quality. Moreover, it can save GPU memory in comparison with the existing works. The experiments show that the quality of reconstructed images and training speed using our architecture is higher than current works which use the SIREN only. Besides, the GPU memory cost is evidently decreased

Figures (14)

• Figure S1: Our suggested architecture using INR (in this work, SIREN) to compress high-resolution (HR) medical images.
• Figure S2: SRDensenet—all architecture with eight blocks.
• Figure S3: Left column shows the high-resolution slices with a size of 512 $\times$ 512 and the right is the low-resolution slices with a size of 128 $\times$ 128. Low-resolution slices were obtained by applying Lanczos resampling on the high-resolution slices.
• Figure S4: From left, the first image shows the original down-sampled image. The second image illustrates the reconstructed counterpart of the down-sampled image using SIREN. The next image shows the plot of the PSNR while training the network (best PSNR: 45.38) and the last image shows the plot of the loss values during the SIREN training (loss: $1.4805471 \times 10^{-5}$).
• Figure S5: From the left, the first image shows the original down-sampled image. The second image illustrates the reconstructed counterpart of the down-sampled image using SIREN. The next image shows the plot of the PSNR while training the network (best PSNR: 48.29) and the last image shows the plot of the loss values during the SIREN training (loss: 1.0805471 $\times \,10^{-5}$).