Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Task-Specific Data Preparation for Deep Learning to Reconstruct Structures of Interest from Severely Truncated CBCT Data

Yixing Huang, Fuxin Fan, Ahmed Gomaa, Andreas Maier, Rainer Fietkau, Christoph Bert, Florian Putz

TL;DR

Capable of reconstructing structures outside a severely truncated CBCT field of view, the paper introduces a task-specific data preparation strategy that segments structures of interest (SOI) and trains a Pix2pixGAN to learn SOI differences from truncated data while preserving data consistency. In a rib-reconstruction exemplar for image-guided needle biopsy, the method reduces false positives and improves Dice similarity for SOI outside the FOV (0.958 vs. 0.906) compared with conventional training, demonstrating robustness to truncation artifacts. The approach enables targeted SOI recovery from limited CBCT data, potentially broadening clinical CBCT applications such as needle-path planning and other CT tasks where only specific structures are clinically relevant.

Abstract

Cone-beam computed tomography (CBCT) is widely used in interventional surgeries and radiation oncology. Due to the limited size of flat-panel detectors, anatomical structures might be missing outside the limited field-of-view (FOV), which restricts the clinical applications of CBCT systems. Recently, deep learning methods have been proposed to extend the FOV for multi-slice CT systems. However, in mobile CBCT system with a smaller FOV size, projection data is severely truncated and it is challenging for a network to restore all missing structures outside the FOV. In some applications, only certain structures outside the FOV are of interest, e.g., ribs in needle path planning for liver/lung cancer diagnosis. Therefore, a task-specific data preparation method is proposed in this work, which automatically let the network focus on structures of interest instead of all the structures. Our preliminary experiment shows that Pix2pixGAN with a conventional training has the risk to reconstruct false positive and false negative rib structures from severely truncated CBCT data, whereas Pix2pixGAN with the proposed task-specific training can reconstruct all the ribs reliably. The proposed method is promising to empower CBCT with more clinical applications.

Task-Specific Data Preparation for Deep Learning to Reconstruct Structures of Interest from Severely Truncated CBCT Data

TL;DR

Capable of reconstructing structures outside a severely truncated CBCT field of view, the paper introduces a task-specific data preparation strategy that segments structures of interest (SOI) and trains a Pix2pixGAN to learn SOI differences from truncated data while preserving data consistency. In a rib-reconstruction exemplar for image-guided needle biopsy, the method reduces false positives and improves Dice similarity for SOI outside the FOV (0.958 vs. 0.906) compared with conventional training, demonstrating robustness to truncation artifacts. The approach enables targeted SOI recovery from limited CBCT data, potentially broadening clinical CBCT applications such as needle-path planning and other CT tasks where only specific structures are clinically relevant.

Abstract

Cone-beam computed tomography (CBCT) is widely used in interventional surgeries and radiation oncology. Due to the limited size of flat-panel detectors, anatomical structures might be missing outside the limited field-of-view (FOV), which restricts the clinical applications of CBCT systems. Recently, deep learning methods have been proposed to extend the FOV for multi-slice CT systems. However, in mobile CBCT system with a smaller FOV size, projection data is severely truncated and it is challenging for a network to restore all missing structures outside the FOV. In some applications, only certain structures outside the FOV are of interest, e.g., ribs in needle path planning for liver/lung cancer diagnosis. Therefore, a task-specific data preparation method is proposed in this work, which automatically let the network focus on structures of interest instead of all the structures. Our preliminary experiment shows that Pix2pixGAN with a conventional training has the risk to reconstruct false positive and false negative rib structures from severely truncated CBCT data, whereas Pix2pixGAN with the proposed task-specific training can reconstruct all the ribs reliably. The proposed method is promising to empower CBCT with more clinical applications.
Paper Structure (9 sections, 2 equations, 3 figures, 1 table)

This paper contains 9 sections, 2 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Task-Specific Data Preparation
  4. Conventional data preparation for FOV extension
  5. Task-specific data preparation for FOV extension
  6. Exemplification with rib reconstruction in image-guided needle biopsy
  7. Experimental Setup
  8. Results And Discussion
  9. Conclusion

Figures (3)

  • Figure 1: Intermediate reconstructions under different data preparations. (a) is the reference slice with all structures. (b) is the input reconstruction from truncated data. (c) and (d) contain ribs and other anatomical structures, respectively. (e) and (f) are truncated reconstruction for ribs and other anatomical structures, respectively. (g) shows the label by conventional data preparation. (h) is the label by task-specific data preparation.
  • Figure 2: The projections of one patient's ribs solely (a) and other anatomical structures (b), where the dashed lines indicate the truncation boundaries, display windows [0, 4.5] for (a) and [0, 9] for (b).
  • Figure 3: A potential application to rib reconstruction from severely truncated data for image-guided needle biopsy, where the structures inside the FOV (indicated by the dashed circle) and the ribs outside the FOV are of interest, window: [-600, 600] HU. (b) is the input image. (c) is the Pix2pixGAN output with conventional data preparation, where a false positive rib indicated by the red arrow is reconstructed and a rib indicated by the blue arrow has a low contrast and incorrect shape. (d) is the Pix2pixGAN output with our proposed task-specific data preparation, where the number of ribs is correct and the positions of ribs are accurate. The two green dash lines are two potential needle paths for the target (marked by the green point) between ribs. Its error image with respective the task-specific label (e) is displayed in (f).