Question about training process

[dragonylee]

Thank you for your work. After reading the paper, I have a question regarding the training process that I would like to ask.

You had mentioned in your paper “For supervision, ground-truth intensity values can be obtained from the ground-truth CT image based on the coordinates of points by trilinear interpolation.” This implies that 3D CT data is needed as supervision during network training. Doesn’t this contradict the concept of "sparse view 3D reconstruction"?

I’m not sure if I misunderstood something, but I hope to get clarification. Thank you!

[lyqun]

3D supervision is necessary during the training as the method is data-driven. However, the 3D is not needed during the inference, and the network can reconstruct the 3D from sparse-view inputs.

[he0713]

Why is a CT reconstruction performed when there is already ground-truth 3DCT data? The reconstructed CT data is certainly not as good as the ground-truth one, so why do you have to do extra work？

[lyqun]

Why is a CT reconstruction performed when there is already ground-truth 3DCT data? The reconstructed CT data is certainly not as good as the ground-truth one, so why do you have to do extra work？

In this setting, CT data is required for supervision during training, which ensures the model learns correct information. During the model inference (testing), the CT is not required, and the inputs are only the sparse projections.

[he0713]

Thanks for the reply, I would also like to know where the ground-truth CT data is obtained in the code?

[lyqun]

Thanks for the reply, I would also like to know where the ground-truth CT data is obtained in the code?

The data was preprocessed and the CT data was saved in blocks for faster data loading. See ./data/knee_cbct/3_block_save.py#L50

In the data loading, the CT values are loaded from a randomly selected block. See ./code/dataset.py#L198

[he0713]

Thanks for your explanation.

…

On Mon, Dec 30, 2024 at 9:05 PM Yiqun Lin ***@***.***> wrote: Thanks for the reply, I would also like to know where the ground-truth CT data is obtained in the code? The data was preprocessed and the CT data was saved in blocks for faster data loading. See ./data/knee_cbct/3_block_save.py#L50 In the data loading, the CT values are loaded from a randomly selected block. See ./code/dataset.py#L198 — Reply to this email directly, view it on GitHub <#8 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BMPEQKA2UKBZR6BG2IGK5CT2IFAJHAVCNFSM6AAAAABQF4S66GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNRVGQ2TQOJVGE> . You are receiving this because you commented.Message ID: ***@***.***>