MRI classification using deeplearning

[Moslem Asgari]

Dear HCP Users,

I am writing to seek expert advice on a deep learning project involving MRI image classification. Given the esteemed expertise in MRI image interpretation within this group, I believe your insights would be invaluable.

I am working with an MRI dataset comprising 16 image slices per subject, divided into 'sick' and 'healthy' categories.Importantly, the MRI modality (T2, T1, or FLAIR) varies across subjects, regardless of health status.

My goal is to develop an accurate deep learning algorithm capable of classifying MRI images as either 'sick' or 'healthy.' In light of this, I would like to pose two questions:

1. Considering the heterogeneity of MRI modalities, would it be advantageous to train separate models for each modality, or is a unified model capable of handling all modalities more suitable?

2. Should the model process all 16 slices as a tensor, or would focusing on a single slice (e.g., the deepest) be more effective?

I eagerly anticipate your valuable perspectives and suggestions. Thank you for your time and consideration.

[Yubo Wang]

Hello,

For question 1, I think taking everything together is a good idea if the model is strong enough to learn the hidden features between them. Although the features (modality) are different, they still have the similar spatial structure which conv kernel will the conv kernel should apply.

For question 2, does "slice" mean to slice 2D image from a 3D volume? If so, is there any reason that you don't want to learn directly from the volume? If there are limited resource to train the network, using some techniques to merge the slices should be better than only taking one slice. But if you want to do it in 2D, I think taking all 16 slices together with some appropriate way to establish the connection between them is a good idea since taking them separately will lose some information about the spatial structure.