應用深度學習降噪暨影像三維重建分析

Abstract

X光激發光學與電腦斷層影像系統結合了X光激發光學和電腦斷層成像技術，為醫學和工業領域提供了一種非常有用的影像診斷工具，提供重要的影像診斷和分析資訊。其利用 X 光對物體進行成像，並同時檢測物體中的激發光信號，具有非侵入性的優點。然而受限於康普敦散射背景雜訊及感測器之熱雜訊，影像重建之分辨率受到限制。隨著深度學習演算法的成熟及電腦硬體的快速發展，近年來圖像降噪技術的性能得到了顯著進展。本研究建立一套 X 光激發光學與電腦斷層影像系統，撰寫反濾波投影及迭代重建演算法來對樣品進行影像三維重建，引入二值池化降噪減少感測器之熱雜訊，並透過深度學習 CCE-3D 模型來改善感測器因熱能、光量子的漲落、暗電流而產生的散粒雜訊，以及因康普敦散射而產生的背景雜訊，最後透過設計樣品及實驗來量化分析影像重建的品質，以Dice 相似性系数（DSC）、結構相似性係數（SSIM）、峰值信噪比（PSNR）等影像分析指標來探討影像三維重建的空間解析度，結果顯示我們成功提升影像重建品質，並將空間解析度提升至0.5 mm。

The X-ray luminescence computed tomography (XLCT) system combines X-ray excitation optics with computerized tomography imaging technology, presenting a highly valuable image diagnostic tool for the medical and industrial domains. The integration facilitates the acquisition of vital image diagnosis and analysis information. By utilizing X-rays for object imaging while concurrently detecting excitation luminescence signals within the object, the XLCT system offers the advantage of non-invasiveness. However, limitations in image reconstruction resolution arise due to background noise from Compton scattering and thermal noise from the sensor. With the advancement of deep learning algorithms and the rapid progress of computer hardware, image-denoising technology has achieved significant development in recent years. This study establishes an X-ray excitation optics and CT imaging system, implementing a filter back projection(FBP) algorithm and simultaneous algebraic reconstruction technique (SART) for three -dimensional prosthesis image reconstruction. To mitigate sensor thermal noise, binary pooling noise reduction techniques are employed. Additionally, the deep learning CCE 3D model is utilized to suppress shot noise, dark current noise, and background noise resulting from Compton scattering. Furthermore, the design of prostheses and corresponding experiments enables quantitative analysis of image reconstruction quality, employing evaluation indexes such as the Dice similarity coefficient, structural similarity coefficient, and peak signal-to-noise ratio to assess the spatial resolution of three- dimensional image reconstruction. The results demonstrate the successful improvement of image reconstruction quality, with the achieved spatial resolution reaching 0.5 mm.