Academic Symposium: Short presentations & panel discussion

Author: Armin Sherbanifard

Title: Representing 3D medical images with Implicit Neural Networks

Abstract: Storing medical images is a challenge in the field of medical imaging. Some methods use Implicit Neural Representation (INR) to compress volumetric medical images, but there’s room for better compression. Many INR techniques require substantial GPU memory and lengthy training for high-quality rendering. In our work, we introduce a novel INR approach that leverages the Lanczos downsampling scheme, SIREN deep network, and SRDenseNet high-resolution scheme. Our architecture reduces training time, achieves high compression rates, and maintains rendering quality. It’s also more memory-efficient. Experiments demonstrate superior image quality, faster training, and reduced GPU memory usage compared to using SIREN alone.