Learning about CT Reconstruction（CT 重建原理 · 上手练习）
Learning about CT Reconstruction (Hands-on Workshop) - User Manual
Dragonfly Prototype Apps · Learning about CT Reconstruction...
版本 Version 1.0 · 2026-07-26
第一部分 中文手册
目录
1. 简介
2. 三个标签页
3. 会讲清楚的原理
4. 最直观的一课
5. 使用步骤
6. 环境需求
1. 简介
Learning about CT Reconstruction（CT 重建原理） 是一个用最直观的方式讲清楚 CT（计算机断层成像）原理与图像重建原理的上手练习项目：为什么绕着物体转一圈、拍很多张 X 光片，就能算出它内部的三维结构。
本功能由第三方（Prototype Labs / Prototype Apps）提供，非 Dragonfly 官方开发或支持，使用风险自负。证书页生成的完成证书亦非官方认证，仅供个人记录。
2. 三个标签页
1. 物体 / 体模 —— 我们要“扫描”的对象是一张衰减分布。一键生成合成体模（棒/球/环三种多材料体模），或加载并体素化 STEP/STL/OBJ/PLY 网格，或直接调用“创建合成 CAD 模型（STEP）”插件自己设计一个，也可选用已有的 Channel。
2. 前向投影（采集） —— 设置 CT 空间参数（投影数 / 角度范围 / 噪声），调用 Radon 变换沿射线积分，生成投影图像（正弦图 sinogram）并发布。这就是真实 CT 采集到的原始数据。
3. 重建 —— 从投影反算断层：滤波反投影（FBP，可选 ramp / shepp-logan / cosine / hamming / hann / none）或迭代重建（SART），并与体模真值对比 RMSE。
3. 会讲清楚的原理
- 成像：X 射线衰减与 Beer–Lambert 定律，投影记录的是沿射线的线积分。
- 采集：正弦图（Radon 变换）、平行/扇/锥束几何、投影数与角度范围的影响。
- 重建：反投影为何必须先滤波、FBP 各滤波器的锐利/噪声权衡、迭代 SART 的取舍、以及欠采样条纹/有限角/射束硬化等常见伪影。
4. 最直观的一课
在“前向投影”页把投影数调很小再重建，能看到明显的放射状欠采样条纹；调大投影数，条纹消失。把滤波器设为 none（纯反投影）会很糊，切回 ramp 立刻清晰。这两组对比直观回答了“为什么 CT 要拍很多张、为什么要滤波”。
5. 使用步骤
1. 在“物体/体模”页点“生成合成体模”（或体素化一个 CAD 网格 / 选一个 Channel）。
2. 在“前向投影”页设好投影数、角度范围（平行束 180° 即足够），点“生成投影”，发布投影与正弦图。
3. 在“重建”页选方法（FBP / SART）与滤波器，点“重建”，查看重建体与 RMSE。
4. 试试少角度 vs 多角度、ramp vs none，观察画质与 RMSE 的变化。
5. 完成后在“证书”页填写姓名生成完成证书。
6. 环境需求
无需安装、无需联网、无需 GPU、无需虚拟环境。核心动手环节用 Dragonfly 自带的 numpy/scikit-image/trimesh 在进程内运行（平行束几何，逐层用 Radon 变换）。加载 STEP 需要 Dragonfly 2027.1 及以上（STL/OBJ/PLY 各版本均可）。
Part II English Manual
Contents
1. Introduction
2. The three tabs
3. The principles it teaches
4. The most vivid lesson
5. How to use
6. Requirements
1. Introduction
Learning about CT Reconstruction is a hands-on project that explains, as intuitively as possible, how CT (computed tomography) works and how image reconstruction works: why rotating around an object and taking many X-ray images lets you compute its internal 3D structure.
This feature is provided by a third party (Prototype Labs / Apps), not officially developed or supported by Dragonfly. The completion certificate is also unofficial — for personal record only.
2. The three tabs
1. Object / phantom — the thing we 'scan' is an attenuation map. Create a synthetic phantom (rods/spheres/rings multi-material phantoms), load and voxelise a STEP/STL/OBJ/PLY mesh, open the 'Create Synthetic CAD Models (STEP)' plugin to design your own, or use an existing Channel.
2. Forward projection (acquisition) — set the CT parameters (number of projections / angular range / noise), run the Radon transform (line integrals along the rays) to produce projection images (a sinogram) and publish them. This is the raw data a real CT records.
3. Reconstruction — recover the cross-sections: filtered back-projection (FBP, filters ramp / shepp-logan / cosine / hamming / hann / none) or iterative SART, compared to the ground-truth object via RMSE.
3. The principles it teaches
- Imaging: X-ray attenuation and the Beer–Lambert law; a projection records a line integral along each ray.
- Acquisition: the sinogram (Radon transform), parallel/fan/cone geometry, and the effect of the number of angles and the angular range.
- Reconstruction: why back-projection must be filtered, the sharpness/noise trade-off of the FBP filters, the pros/cons of iterative SART, and common artifacts (undersampling streaks, limited angle, beam hardening).
4. The most vivid lesson
On the 'Forward projection' tab, set a small number of projections and reconstruct — you see clear radial undersampling streaks; raise the number and they vanish. Set the filter to none (plain back-projection) and it is very blurry; switch back to ramp and it is instantly sharp. These two comparisons answer, viscerally, 'why CT takes so many images' and 'why you must filter'.
5. How to use
1. On 'Object / phantom' click 'Create synthetic phantom' (or voxelise a CAD mesh / pick a Channel).
2. On 'Forward projection' set the number of projections and range (180° suffices for parallel beam), click 'Generate projections', and publish the projections + sinogram.
3. On 'Reconstruction' choose the method (FBP / SART) and filter, click 'Reconstruct', and read the volume + RMSE.
4. Try few vs many angles, and ramp vs none, to watch the quality and RMSE change.
5. Finish on the 'Certificate' tab by entering your name to generate a completion certificate.
6. Requirements
No install, no internet, no GPU, no venv. The core actions run in-process on Dragonfly's own numpy/scikit-image/trimesh (parallel-beam geometry, slice-by-slice Radon transform). Loading STEP needs Dragonfly 2027.1+ (STL/OBJ/PLY work on all versions).