search

konar1987 / 3D-QNet

3D-QNet for Volumetric Medical Image Segmentation
Boost Software License 1.0
0 stars 0 forks source link

Choice of Cluster values #1

Open laiadc opened 7 months ago

laiadc commented 7 months ago

In the Transit function, you define an array called 'cluster', for different lenghts (L=4, 6, 7, 8). It seems to me that the choice of this cluster array highly influences the output and performance of the algorithm. I understand it is a hyperparameter of the algorithm. Do you have any idea of what is the best way to choose this array, based on the input image? Could you describe what the role of this cluster variable?

Thank you so much!

Best,

Laia.

konar1987 commented 7 months ago

Dear Laia,

Thanks for your interest in our 3D-QNet.

The clusters are adaptive and based on the optimization on input images. The cluster decises the boundaries for multi-level sigmoid activation function which we have proposed.

If you have any further questions, please let me know in my mail id: @.***

Best regards,

Debanjan

On Mar 11, 2024, at 10:13 AM, Laia Domingo Colomer @.***> wrote:

In the Transit function, you define an array called 'cluster', for different lenghts (L=4, 6, 7, 8). It seems to me that the choice of this cluster array highly influences the output and performance of the algorithm. I understand it is a hyperparameter of the algorithm. Do you have any idea of what is the best way to choose this array, based on the input image? Could you describe what the role of this cluster variable?

Thank you so much!

Best,

Laia.

— Reply to this email directly, view it on GitHub https://github.com/konar1987/3D-QNet/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AR4YPQHM5LGHDRV6VEUBUDLYXW3Y7AVCNFSM6AAAAABEQMQ4FGVHI2DSMVQWIX3LMV43ASLTON2WKOZSGE3TSMRUGEYDAMI. You are receiving this because you are subscribed to this thread.

darkEnvious commented 1 month ago

Hi,

Firstly, thanks for making your codebase public - it is very helpful to browse through the code while reading your paper on 3-D Quantum-Inspired Self-Supervised Tensor Network for Volumetric Segmentation of Medical Images.

I had a question on the above lines - what are the parameters in the code that get updated according to the learning process? Or is the provided code applying pre-defined Quantum-Inspired transformations and then using classical segmentation models? I was just curious and wanted to have a look at how this model achieves self-supervised learning.

Thanks in advance, and I am looking forward to hearing from you!

Regards, Nirmal

konar1987 commented 1 month ago

Dear Nirmal,

Thanks for your interest in our work.

The weight parameters are updated in a self-supervised fashion without any training. We have proposed an intelligent 3D decomposition of weights (Tucker Decomposition) method in the self-learning process.

Please let me know if you have any further questions.

Best regards,

Debanjan

On Wed, Sep 11, 2024 at 1:05 PM Nirmal Sethumadavan < @.***> wrote:

Hi,

Firstly, thanks for making your codebase public - it is very helpful to browse through the code while reading your paper on 3-D Quantum-Inspired Self-Supervised Tensor Network for Volumetric Segmentation of Medical Images.

I had a question on the above lines - what are the parameters in the code that get updated according to the learning process? Or is the provided code applying pre-defined Quantum-Inspired transformations and then using classical segmentation models? I was just curious and wanted to have a look at how this model achieves self-supervised learning.

Thanks in advance, and I am looking forward to hearing from you!

Regards, Nirmal

— Reply to this email directly, view it on GitHub https://github.com/konar1987/3D-QNet/issues/1#issuecomment-2342882613, or unsubscribe https://github.com/notifications/unsubscribe-auth/AR4YPQC3EUFDQLKUF7JCOVTZV7XENAVCNFSM6AAAAABOAKKXN2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNBSHA4DENRRGM . You are receiving this because you commented.Message ID: @.***>

darkEnvious commented 1 month ago

Hi, thanks a lot for your response!

I understand that the weight parameters are updated through an iterative process, as described in the Transit.m file. However, is the Tucker Decomposition method included in this code, or is it only proposed in the paper? I am curious to look at the differences between the non-tensor 3D-QNet and the tensor-powered 3D-QNet first-hand, and I would appreciate your insights here!

Thanks in advance!

Regards, Nirmal