Script for VAE evaluation

[matt-baugh]

Please could you provide the script for your VAE evaluation (measuring the LPIPS, SSIM, PSNR provided in table 2 of the README.md?

I tried to make my own script using the VAE for reconstruction but the peak GP memory usage was much higher than you reported, so I'm not sure if there is something wrong with my implementation or maybe there is a difference in my environment causing the issue.

Either way, it would be great to see the original code to keep everything nice and comparable.

[Can-Zhao]

If the GPU memory is high, you could reduce sliding window size "val_sliding_window_patch_size" in configs/config_maisi_vae_train.json

Validation demo is included in https://github.com/NVIDIA-Medtech/NV-Generate-CTMR/blob/main/train_vae_tutorial.ipynb . Is the peak GP memory usage perform as expected?

Please use dynamic_infer instead of direct inference. It will automatically choose between sliding window inference and direct inference depending on the image size.

val_sliding_window_patch_size = args.val_sliding_window_patch_size
val_inferer = SlidingWindowInferer(
        roi_size=val_sliding_window_patch_size,
        sw_batch_size=1,
        progress=False,
        overlap=0.0,
        device=torch.device("cpu"),
        sw_device=device,
    )
reconstruction, z_mu, z_sigma = dynamic_infer(val_inferer, autoencoder, images)

[paproaskwar]

When using a pre-trained VAE to reconstruct CT images, the reconstructed HU values are also in the 0–1 range due to normalization. Can these values be restored to be consistent with the HU values of the original CT images?

[Can-Zhao]

When using a pre-trained VAE to reconstruct CT images, the reconstructed HU values are also in the 0–1 range due to normalization. Can these values be restored to be consistent with the HU values of the original CT images?

Yes, you can simply linear map 0--1 to -1000 -- 1000

[paproaskwar]

I have already tried this method, but there is still a significant HU deviation between the reconstructed CT and the original CT

…

---- Replied Message ---- | From | Can ***@***.***> | | Date | 04/13/2026 22:32 | | To | NVIDIA-Medtech/NV-Generate-CTMR ***@***.***> | | Cc | paproaskwar ***@***.***>, Comment ***@***.***> | | Subject | Re: [NVIDIA-Medtech/NV-Generate-CTMR] Script for VAE evaluation (Issue #2) | Can-Zhao left a comment (NVIDIA-Medtech/NV-Generate-CTMR#2) When using a pre-trained VAE to reconstruct CT images, the reconstructed HU values are also in the 0–1 range due to normalization. Can these values be restored to be consistent with the HU values of the original CT images? Yes, you can simply linear map 0--1 to -1000 -- 1000 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Can-Zhao]

I have already tried this method, but there is still a significant HU deviation between the reconstructed CT and the original CT
…

May I ask if your image is head CT? It is a known issue that the HU cannot keep well for brain region inside head CT.

[paproaskwar]

No…it's lung CT

…

---- Replied Message ---- | From | Can ***@***.***> | | Date | 04/14/2026 04:49 | | To | ***@***.***> | | Cc | ***@***.***>***@***.***> | | Subject | Re: [NVIDIA-Medtech/NV-Generate-CTMR] Script for VAE evaluation (Issue #2) | Can-Zhao left a comment (NVIDIA-Medtech/NV-Generate-CTMR#2) I have already tried this method, but there is still a significant HU deviation between the reconstructed CT and the original CT … May I ask if your image is head CT? It is a known issue that the HU cannot keep well for brain region inside head CT. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Can-Zhao]

No…it's lung CT
…

That's unexpected. Is the imaged clipped [-1000,1000] before sending to the VAE? It will nor recover intensity outside this range. For the intensity inside this range, may I ask how much off you have observed?

[paproaskwar]

Yes, the CT has been clipped to the range of -1000 to 1000. The following figure displays the HU difference along the same line between the two CTs. 在 2026-04-14 10:55:59，"Can Zhao" ***@***.***> 写道： Can-Zhao left a comment (NVIDIA-Medtech/NV-Generate-CTMR#2) No…it's lung CT … That's unexpected. Is the imaged clipped [-1000,1000] before sending to the VAE? It will nor recover intensity outside this range. For the intensity inside this range, may I ask how much off you have observed? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Can-Zhao]

Yes, the CT has been clipped to the range of -1000 to 1000. The following figure displays the HU difference along the same line between the two CTs.

在 2026-04-14 10:55:59，"Can Zhao" @.***> 写道：

Can-Zhao left a comment (NVIDIA-Medtech/NV-Generate-CTMR#2)

No…it's lung CT
…

That's unexpected. Is the imaged clipped [-1000,1000] before sending to the VAE? It will nor recover intensity outside this range. For the intensity inside this range, may I ask how much off you have observed?

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you commented.Message ID: @.***>

It seems the figure was not uploaded..