Hello All,
I have been through many research papers where I have seen that in deep-learning applications, people use only a few slices (some 3 or some 5) for the MRI data-based diagnosis. But the actual MRI files contains a lot of slices in both ADNI or OASIS dataset. Can someone help me how to choose these slices for the deep learning applications.

I think you can choose all slices and do some convolution to reduce slices number

Hi,
I have tried taking all slices at once, but the the accuracy obtained in this case is very low. 
Can you suggest any Git repository or any published work where such work is being done?

Thanks in advance

Accuracy will be low since unwanted region will also contribute to feature learning. Selective slices is the only approach. Selection of slice can be done through several algorithm 

Thanks Rahul,

Can you suggest me some algorithm by which I can select the important slice that will contribute to increase the accuarcy. because it is very important which slice to consider.

"Selection of slice can be done through several algorithm"

Please point to any of those algorithms. 

Thanks

I can recomand you some related papers which focus on using attention method to select important for all MRI slices.Interpretable medical deep framework by logits-constraint attention guiding graph-based multi-scale fusion for Alzheimer’s disease analysis and Cross-attention guided loss-based deep dual-branch fusion network for liver tumor classification these two papers focus on this.You can check them.

Based on the literature I have seen, I believe it's most common to select slices from the central regions. I personally have used all sagittal slices so far, but I am planning to use focus on the slices near the central sagittal plane.

You can refer to these papers and check if the method suits your study ->

An analysis of data leakage and generalizability in MRI based classification of Parkinson's Disease using explainable 2D Convolutional Neural Networks (doi: https://doi.org/10.1016/j.dsp.2024.104407)

Effect of data leakage in brain MRI classification using 2D convolutional neural networks (doi: https://doi.org/10.1038/s41598-021-01681-w)

Deep learning-based approach for Parkinson’s disease detection using region of interest (doi: https://doi.org/10.1007/978-981-16-2422-3_1)