Denoising study of Fluoroscopic Images in real time tumor tracking System based on Statistical model of noise

TL;DR

This paper analyzes the noise characteristics of intraoperative fluoroscopic images in real-time tumor tracking and develops a noise model to improve denoising performance using deep learning.

Contribution

It introduces a novel noise generation method based on statistical noise patterns and demonstrates improved denoising results with a deep learning model trained on this data.

Findings

Proposed noise model better captures real noise patterns.

Deep learning model trained on the proposed noise data outperforms Gaussian noise-based training.

Enhanced image quality improves tumor tracking accuracy.

Abstract

This study investigates the noise characteristics of intraoperative X-ray fluoroscopic images acquired during real-time image-guided radiotherapy (IGRT), and presents a novel noise image generation method based on the identified noise amplitude and spatial probability patterns. Initially, noise-free digitally reconstructed radiographs (DRRs) were generated using patient CT data combined with projection algorithms and the spatial configuration of the real-time tumor tracking system. Based on the observed noise probability and amplitude distributions, noise was then added to these DRRs to create Dataset 1. As a control, Dataset 2 was generated by adding Gaussian noise with the same mean and variance as Dataset 1; however, the noise probability in Dataset 2 is independent of pixel location and pixel intensity. Both datasets were used to fine-tune a pre-trained SwinIR model with identical training parameters. Tests on phantom images containing real noise show that the SwinIR model trained with the proposed noise model dataset achieves superior denoising performance over the model trained with Gaussian noise and the model without transfer learning, with an average PSNR improvement of 1.45 dB. This study contributes to a deeper understanding of noise patterns in these fluoroscopic images and is crucial for enhancing image quality and the accuracy of real-time tumor tracking in radiotherapy.