Experimental design
This protocol is extracted from research article:
Respiratory deformation registration in 4D-CT/cone beam CT using deep learning
Quant Imaging Med Surg, Feb 1, 2021; DOI: 10.21037/qims-19-1058

The goal of the network is to predict the DVF that registers the input image pair efficiently. In order to test the generalizability of our network under different circumstances, different studies combining different training and testing data sets were used, as shown in Table 1. Overall, nine 4D-CT volumes from eight patients with lung cancer from Duke University Medical Center were used for study 1, 2, 3 and 6. The study was approved by the Ethics board of Duke (No.: 00049474) and informed consent was taken from all the patients. Also, the 4D-CBCT volumes and projections from the public datasets Sparse-view Reconstruction Challenge for Four-dimensional Cone-beam CT (SPARE) (12) were used for study 4 and 5. For each dataset, there are 10 phases for the respiratory motion.

The first five studies are intra-patient studies. The last one is the inter-patient study.

The training and predicting samples are from the same 4D-CT volume. The network is trained with learning the deformation from phase 1 to phase 6, and it is applied to predict the deformation from phase 1 to phase 7. The feasibility of CNN for deformable registration was tested with this study.

The training and predicting samples are from the two 4D-CT volumes from the same patient with lung cancer. The model is trained to learn the deformation from phase 1 to all the other phases from the first set of 4D-CT volume, and it is applied to predict the deformation from phase 1 to all the other phases on the second 4D-CT volume. Two sets of 4D-CT volumes were taken at different times. Therefore anatomic difference was introduced between training and predicting samples to test the network’s robustness against inter-fractional anatomical differences.

Two simulated 4D-CBCT were reconstructed from digital reconstructed radiography (DRRs) generated from the double 4D-CT in study 2 based on the cone-beam geometry. The training and predicting setup is the same as trial 2. The sampling artifacts of cone-beam geometry was introduced in this test.

Two 4D-CBCT volumes were reconstructed with FDK methods using the projections from SPARE. The training volume is reconstructed with 680 projections/phase for high image quality, and the predicting volume is reconstructed with 170 projections/phase to mimic the clinic 4D-CBCT acquisition. The adaptability of the network to poor image quality in predicting volume was tested.

A ground truth 4D-CT volume and an FDK-reconstructed 4D-CBCT volume from Sparse-view Reconstruction Challenge for Four-dimensional Cone-beam CT (SPARE) (12) were used. The ground truth 4D-CT volume is used for training, and the FDK-reconstructed 4D-CBCT is used for predicting. There is a tumor in ground truth volume but not in reconstructed 4D-CBCT volume; therefore, the adaptability of the network to the big anatomic difference between training and predicting sets was predicted. In addition, the robustness of the network against cross-modality differences was predicted.

Six 4D-CT volumes from six patients with lung cancer from Duke University Medical Center were used. Five of them were used as the training samples, and the last one was used for predicting. The interpatient anatomic difference between training and testing volumes was introduced.

Note: The content above has been extracted from a research article, so it may not display correctly.



Q&A
Please log in to submit your questions online.
Your question will be posted on the Bio-101 website. We will send your questions to the authors of this protocol and Bio-protocol community members who are experienced with this method. you will be informed using the email address associated with your Bio-protocol account.



We use cookies on this site to enhance your user experience. By using our website, you are agreeing to allow the storage of cookies on your computer.