Deep Learning CT Reconstruction From Incomplete Projection Data

Original link: A Review of Deep Learning CT Reconstruction From Incomplete Projection Data

---

📄 Abstract

This article provides a comprehensive overview of the current state-of-the-art DL-based CT Reconstruction from incomplete projection data, including 5 problems:

1. Sparse view reconstruction
2. Limited angle reconstruction
3. Metal artifact reduction
4. Interior tomography
5. Ring artifact reduction

---

🔬 Ⅰ. Introduction

The quality of the reconstructed images from CT scans is directly influenced by the completeness of the measured data. Accurate CT images can be reconstructed from complete projection data using classic analytical or iterative reconstruction methods. However, complete projection data may not be available in practice, due to factors such as radiation dose limitations, space and time constraints, object shape, and imaging system limitations.

Before deep learning, early methods for incomplete projection data reconstruction fell into two categories:

1. Projection data completion methods — start from the projection data and use known information to estimate missing data
2. Iterative reconstruction methods — formulate an objective function based on observation and prior knowledge of CT images, then optimize

In recent years, a wide variety of DL-based methods have emerged and brought new solutions to incomplete projection data reconstruction.

---

🩻 Ⅱ. Causes of Incomplete Projection Data

In parallel-beam imaging systems:

• Full scan: 360° scanning angle
• Half scan: 180° — the minimum angle required for accurate reconstruction

In fan-beam imaging systems, the half-scanning range is 180 + w°, where w is the field angle of the fan beam.

Note: This survey uses the fan-beam scanning system as the primary example, since fan-beam is more realistic and more commonly studied in clinical practice.

A. Incomplete Sampling View Problem

Two sub-problems:

Sparse View (SV) Sampling — only part of the projection data within the full scan range is obtained. Used to reduce radiation exposure. Also shortens scanning time and reduces motion artifacts from heartbeat and respiratory motion. Side effect: streak artifacts introduced into the reconstructed image.

Limited Angle (LA) Sampling — similar data incompleteness, but due to angular range limitations rather than angular density.

Warning: Streak artifacts from sparse-view sampling can impair important clinical details. The goal of DL methods is to suppress these artifacts while preserving diagnostic information.

B. Metal Implant Problem

Content coming soon.

C. Interior Problem

Content coming soon.

D. Ring Artifact Problem

Content coming soon.

---

🤖 Ⅲ. Review of Current Deep-Learning-Based Methods

A. Different Domain Learning

DL-based methods can be classified into four distinct groups based on the data domains they operate in:

flowchart TD
    DL["🤖 DL-Based CT Reconstruction"]:::root
    DL --> SIN["📡 Sinogram Domain\nLearning"]:::dom
    DL --> IMG["🖼️ Image Domain\nLearning"]:::dom
    DL --> DUAL["🔁 Dual-Domain\nLearning"]:::dom
    DL --> TRANS["🔄 Domain Transformation\nLearning"]:::dom

    classDef root fill:#4A90D9,stroke:#2c5f8a,color:#fff
    classDef dom fill:#5BA85A,stroke:#3a6e39,color:#fff

Tip: Dual-domain networks are integrated to effectively extract information from both sinogram and image domains and complement each other's weaknesses.

B. Unsupervised and Semi-Supervised Learning

Content coming soon.

---

📊 Ⅳ. Evaluation

A. Dataset

1. DeepLesion
2. Spineweb Dataset 3
3. AAPM
4. CTPelvic1K

B. Deep Reconstruction Tools

C. Artifacts Reduction Performance

Content coming soon.

D. Stability Performance

Content coming soon.

---

🔮 Ⅴ. Conclusion and Future Works

Content coming soon.

---

Part of my CT reconstruction paper review series. Next: specific DL architectures for sinogram-domain artifact reduction.