Joint Denoising and 3D Point Cloud Reconstruction from Single Medical Images

NLDL 2026 Conference Submission68 Authors

Keywords: Medical Image Denoising, 3D Reconstruction, Point Cloud, Joint Learning, Deep Learning

Abstract: We present a preliminary joint learning framework for simultaneous image denoising and 3D point cloud reconstruction from single noisy medical images. The model employs a two-branch architecture with shared intermediate representations, offering a compact alternative to sequential pipelines in resource-constrained environments. Due to hardware and dataset access limitations, we conducted proof-of-concept experiments on synthetic noisy data. Results yield a PSNR of approximately 10 dB, SSIM of 0.34, and Chamfer Distance of 0.05 (mean ± std across seeds). While these numbers are modest, they demonstrate the feasibility of coupling denoising and reconstruction within a single model. We outline challenges, including reliance on synthetic data and limited GPU memory, and discuss future directions toward real LIDC-IDRI validation, efficiency benchmarking, and integration with recent diffusion- and transformer-based methods. This study provides an early step toward compact multi-task models for clinical imaging workflows.

Serve As Reviewer: ~Ming-Hui_Liu1

Submission Number: 68