Need some explanations on saver's output?

[tctco]

Could you provide a little more detailed explanation of the model's outputs? I'm not sure if I got everything right :(

• content_input*.nii.gz: 'content' of an image, not related to classification
• gen_clc*.nii.gz / gen_recon*.nii.gz: image being translated to the other class (?), they should be very similar, eg healthy image -> reconstructed healthy image & AD image -> reconstructed AD image
• gen_fake*.nii.gz / gen_random*.nii.gz: image being translated to the other class, but using different translating strategy
• gen_diff*.nii.gz: feature maps

Another question is: suppose I have a patient's image, and I wanna see how ICAM interprets this image. It seems that I have to also give ICAM a healthy image so that ICAM can conduct the translation, is that right? Is it possible to give ICAM only the patient's image without a control image during testing/prediction and get the feature map? This problem is important to me since some patients in my scenario are still cognitively normal but have subject memory complaints, and I wanna see if they truly have some imaging alternations compared to healthy volunteers.

Thanks!

[ecr05]

Ivan, can you answer this please? Sent from Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: Tang Cheng ***@***.***> Sent: Wednesday, March 26, 2025 1:00:48 PM To: CherBass/ICAM ***@***.***> Cc: Subscribed ***@***.***> Subject: [CherBass/ICAM] Need some explanations on saver's output? (Issue #9) Could you provide a little more detailed explanation of the model's outputs? I'm not sure if I got everything right :( * content_input*.nii.gz: 'content' of an image, not related to classification * gen_clc*.nii.gz / gen_recon*.nii.gz: image being translated to the other class (?), they should be very similar * gen_fake*.nii.gz / gen_random*.nii.gz: image being translated to the other class, but using different translating strategy * gen_diff*.nii.gz: feature maps Another question is: I have a patient's image, and I wanna see how ICAM interprets this image. It seems that I have to also give ICAM a healthy image so that ICAM can conduct the translation, is that right? Is it possible to give ICAM only the patient's image without a control image during testing/prediction? Thanks! — Reply to this email directly, view it on GitHub<#9>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AD7GXPGP4OEAGYAGORRWLYD2WKQIBAVCNFSM6AAAAABZ2HZNAKVHI2DSMVQWIX3LMV43ASLTON2WKOZSHE2DSNJSHEYTEOA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***> [tctco]tctco created an issue (CherBass/ICAM#9)<#9> Could you provide a little more detailed explanation of the model's outputs? I'm not sure if I got everything right :( * content_input*.nii.gz: 'content' of an image, not related to classification * gen_clc*.nii.gz / gen_recon*.nii.gz: image being translated to the other class (?), they should be very similar * gen_fake*.nii.gz / gen_random*.nii.gz: image being translated to the other class, but using different translating strategy * gen_diff*.nii.gz: feature maps Another question is: I have a patient's image, and I wanna see how ICAM interprets this image. It seems that I have to also give ICAM a healthy image so that ICAM can conduct the translation, is that right? Is it possible to give ICAM only the patient's image without a control image during testing/prediction? Thanks! — Reply to this email directly, view it on GitHub<#9>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AD7GXPGP4OEAGYAGORRWLYD2WKQIBAVCNFSM6AAAAABZ2HZNAKVHI2DSMVQWIX3LMV43ASLTON2WKOZSHE2DSNJSHEYTEOA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[ecr05]

Hello Tang Cheng, Happy to provide a little more detailed explanation of the model's outputs. 1) If we take disentanglement as an example, then the content input should be an image without disease, containing only content information. 2) The gen_clc and gen_recon should be related to the algorithm of cyclegan, which makes sure that the model is capturing the right information. 3) Gen_fake should be translated to the other class, the gen_random should be sampling randomly from the latent space. 4) the gen_diff would be the difference map between the translated image and the original image, so that we can see if the model is actually predicting the right disease information. 5) If you have a patient's image, and you wanna see how ICAM interprets this image, you don't need a healthy image to translate; you only need to interpret on the latent space then you can get what you want. You will only need healthy and unhealthy images during training. Feel free to let me know if you have further questions. Thanks! Best, Zhenshan.

…

________________________________ 发件人: Emma Robinson ***@***.***> 发送时间: 2025年3月26日 13:30 收件人: CherBass/ICAM ***@***.***>; CherBass/ICAM ***@***.***> 抄送: Subscribed ***@***.***>; Zhenshan Xie ***@***.***> 主题: Re: [CherBass/ICAM] Need some explanations on saver's output? (Issue #9) Ivan, can you answer this please? Sent from Outlook for Android<https://aka.ms/AAb9ysg>

________________________________ From: Tang Cheng ***@***.***> Sent: Wednesday, March 26, 2025 1:00:48 PM To: CherBass/ICAM ***@***.***> Cc: Subscribed ***@***.***> Subject: [CherBass/ICAM] Need some explanations on saver's output? (Issue #9) Could you provide a little more detailed explanation of the model's outputs? I'm not sure if I got everything right :( * content_input*.nii.gz: 'content' of an image, not related to classification * gen_clc*.nii.gz / gen_recon*.nii.gz: image being translated to the other class (?), they should be very similar * gen_fake*.nii.gz / gen_random*.nii.gz: image being translated to the other class, but using different translating strategy * gen_diff*.nii.gz: feature maps Another question is: I have a patient's image, and I wanna see how ICAM interprets this image. It seems that I have to also give ICAM a healthy image so that ICAM can conduct the translation, is that right? Is it possible to give ICAM only the patient's image without a control image during testing/prediction? Thanks! ― Reply to this email directly, view it on GitHub<#9>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AD7GXPGP4OEAGYAGORRWLYD2WKQIBAVCNFSM6AAAAABZ2HZNAKVHI2DSMVQWIX3LMV43ASLTON2WKOZSHE2DSNJSHEYTEOA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***> [tctco]tctco created an issue (CherBass/ICAM#9)<#9> Could you provide a little more detailed explanation of the model's outputs? I'm not sure if I got everything right :( * content_input*.nii.gz: 'content' of an image, not related to classification * gen_clc*.nii.gz / gen_recon*.nii.gz: image being translated to the other class (?), they should be very similar * gen_fake*.nii.gz / gen_random*.nii.gz: image being translated to the other class, but using different translating strategy * gen_diff*.nii.gz: feature maps Another question is: I have a patient's image, and I wanna see how ICAM interprets this image. It seems that I have to also give ICAM a healthy image so that ICAM can conduct the translation, is that right? Is it possible to give ICAM only the patient's image without a control image during testing/prediction? Thanks! ― Reply to this email directly, view it on GitHub<#9>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AD7GXPGP4OEAGYAGORRWLYD2WKQIBAVCNFSM6AAAAABZ2HZNAKVHI2DSMVQWIX3LMV43ASLTON2WKOZSHE2DSNJSHEYTEOA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[tctco]

Thanks for your elaboration!

However, the 'content' image is quite noisy and is downsampled to only (32 x 40 x 32). The 'faked' or the translated image looks better.

real_input0_class0	content_input0_class0	gen_fake_input0_class1

class 1: patient image; class 0: healthy control

Since I want a high-resolution attribution map, is it possible to directly subtract the converted image from the original image? Here is what I did, and I wonder if this is correct:

for img, label, _ in tqdm(dataloader):
    img = img.to("cuda")
    label[:, 1] = 1 # no matter whether the input is AD or CN, take it as AD and let the model convert it to CN
    label = label.to("cuda")
    with torch.no_grad():
        output, _, diff_AD_minus_CN, *_ = model.test_forward_random_group(
            img, label
        )

According to your code:

# model.py

class ICAM:
...
    def test_forward_random_group(self, image, c_org=None, num=50):
        """
        Method for translation from one image of one class to another class.
        Using rejection sampling - this will give the mean and variance maps.
        :param image: image input
        :param c_org: label of image
        :param num: number of times to sample from attribute latent space (for mean and variance maps)
        :return:
        """
        ...
        return output, diff_m_pos_mean, diff_m_neg_mean, diff_m_pos_std, diff_m_neg_std
...

If this is not the correct way to interpret ICAM output, could you provide a small demo to show how I can get the attribution map correctly? Terribly sorry for the inconvenience :( I'm from the medical side and not very familiar with image translation protocols and your work, though inspiring and solid, looks a little too complicated for me 😢

Thanks!