Compare performance of 2.5d model and 2d model

[sam-may]

Baseline 2d model uses only the single CT slice of interest in predicting pneumonia. A 2.5d model would use the n slices above/below the slice of interest as additional training features.

How to add the additional 2n slices? Some options:

1. Add them as additional channels to the single CT slice of interest. I.e. training input shapes will change from M x M x 1 to M x M x (2n + 1)
2. Have 2n + 1 separate input images (this seems less elegant, and I don't see a real advantage in doing it this way)

How to implement?

1. Create a function analogous to load_dcms https://github.com/sam-may/Covid-19-Pneumonia/blob/master/utils.py#L20-L38 which returns an array of 2n + 1-channel images rather than 1-channel images as is currently done.
2. Other downstream functions in data_helper.py probably will need to be updated as well: https://github.com/sam-may/Covid-19-Pneumonia/blob/master/data_helper.py#L133-L136 https://github.com/sam-may/Covid-19-Pneumonia/blob/master/data_helper.py#L160-L169
3. Update DNN architecture:
   • One option is to just indicate that there are additional channels, i.e. change 1 -> 2n + 1 in this line: https://github.com/sam-may/Covid-19-Pneumonia/blob/master/models.py#L85
   • Might also want to explore fully 3d convolutions. This would entail making analogous 3d versions of these functions: https://github.com/sam-may/Covid-19-Pneumonia/blob/master/models.py#L28-L72 . The first option is probably a good starting point.

How to compare?

1. Quantitative: compare metrics e.g. dice coefficient, cross-entropy, accuracy, etc.
2. Qualitative: compare original image, ground truth, and predictions for 2d/2.5d models

[sam-may]

@jkguiang since we are still waiting on getting you and Amir access to the OSG machine where I have the full (deidentified) sets of CT scans stored, here is one option we could do:

1. I transfer 5-10% of the data to the prp-gpu-1 machine and you can use this to work on the necessary modifications to data prep (e.g. in utils.py and data_helper.py).
2. Once this looks good, you can make a PR and I will run the data prep on the OSG machine and transfer the output hdf5 file to prp-gpu-1.
3. Then you can proceed with working on modifying the architecture and comparing performance with the 2d model.

Edit: I am transferring the Russia cohort of patients (50 patients, but fewer slices/patient and lower resolution than the Wuhan cohort so it's much smaller in size) to /xilinx/scratch/covid_ct_data/russia on prp-gpu-1.