Hello, i am PhD student working on a study aimed at classifying drug-naïve schizophrenia patients into treatment-responsive and treatment-resistant groups using baseline MRI data. As an initial step, I am focusing on the utility of T1-weighted MRI brain images. Specifically, I am using 1,200 T1 MRI images from the Human Connectome Project (HCP) dataset to train a 3D Convolutional Autoencoder (CAE). The objective is to identify a lower-dimensional latent space that best represents the HCP data. This trained model will be utilized for various downstream processes, including:
- Feature extraction for our binary classification problem.
- Construction of reconstruction error maps for analysis on both whole-brain and region-of-interest (ROI) bases.
While we are confident in our machine learning strategy, we have several critical questions regarding MRI image preprocessing steps:
1. Anatomical Regions for Model Training:
- Question: Should we focus exclusively on gray matter (GM) by removing white matter (WM) and cerebrospinal fluid (CSF) spaces, or is it more advantageous to use the entire brain-extracted image?
- Consideration: We are contemplating defining a custom loss function that penalizes reconstruction errors in GM voxels more heavily. Would this approach enhance the model’s sensitivity to GM, which is particularly relevant in schizophrenia research?
2. Optimal Denoising Algorithm:
- Question: We initially experimented with Gaussian denoising but observed significant loss of information due to over-smoothing. We have received recommendations to use Non-Local Means (NLM) denoising.
- Request: Do you have alternative strategies or recommendations for denoising methods that preserve critical anatomical details in T1-weighted images?
3. Image Registration and Voxel Dimensions:
- Question: The HCP dataset has voxel dimensions of 0.7 × 0.7 × 0.7 mm. Should we:
- Use an MNI template with a voxel size of 1 × 1 × 1 mm and resample the template to match the HCP dimensions?
- Resample the HCP data to match a higher-resolution MNI template (e.g., 0.5 mm)?
- Challenge: Our patient dataset for the downstream task has a lower resolution of 1 mm. How should we best address these discrepancies in voxel dimensions to ensure optimal alignment and comparability?
4. Impact of Registration on Image Quality:
- Observation: We noticed significant blurring of the HCP dataset quality when applying affine and non-linear registration methods.
- Question: Would it be advisable to skip registration altogether and rely on the autoencoder model to account for individual variability, thereby preserving maximal information? Alternatively, should we perform at least rigid registration to ensure some degree of uniformity across images?
5. Recommendations for Preprocessing Tools:
- Current Plan: We intend to use ANTsPy for bias field correction, NLM denoising, brain extraction (via ANTsPyNet), and registration.
- Request: Do you have any specific recommendations for other tools or pipelines that might be better suited for our objectives?
Would really appreciate guidance form someone with expertise on MRI Brain preprocessing