Guide.txt

Simple guide to run experiments similar to done for the RT Dose paper.
Broad category of experiments include No Beam, Beam + MAE loss, and Beam + MAE + DVH Loss.
These 3 experiments are done for both Manual and ECHO dose data.

To run No-Beam experiment:
1) File data/base_dataset.py 
      line 221: ptv *= 60
      For the with beam case we set the PTV area in the Beam to 60 but in case of no beam that'll have no effect
      as we don't input the beam. So we set PTV to 60 directly (intead of binary 0/1 it's 0/60 now).
2) File data/dosepred3d_dataset.py
      lines 71/72 for training and lines 79/80 for testing.
      We don't send Beam as an input so, here we can decide which inputs to concatenate to the nework as input.
      Input A is CT and OAR is one-hot encoded OAR/PTV. 5 OAR, 1 PTV and 1 CT make total 7 channel input.
3) File models/doseprediction3d_model.py
      lines 44 & 46; 63 & 65; 100, 102 and 111
      In these groups of lines we select which loss to use MAE or MAE + DVH. Uncomment the lines for DVH loss
      if we want to use the latter. Finally, in line 111 add all the losses together if there are more than 1.
4) File scripts/train_ct2dose3d.sh or test_ct2dose3d.sh
      python train.py --dataroot ./datasets/msk-manual-3d-dvh-beamlet-dense-separate-ptv --netG unet_128 --name EXPERIMENT-NAME --model doseprediction3d --direction AtoB --lambda_L1 1 --dataset_mode dosepred3d --norm batch --batch_size 1 --pool_size 0 --display_port 8097 --lr 0.0002 --input_nc 7 --output_nc 1 --display_freq 10 --print_freq 1 --gpu_ids 0
      
      the "dataroot" points to training data which I have shared with Dr. Nadeem on the google drive for the project. "netG" defines which CNN model we want to train (file models/networks3d.py), we can implement more models here. "input_nc" will be 7 for NO-BEAM case and 8 for with beam case. other options like "display_freq" etc. you can look up what they do in files "options/base_options,train_options/test_options.py).
      defaul display port is 8097 for visdom server. Once you install visdom, start the server and navigate to localhost.ocaldomain:8097 in your browser it will plot the losses which we have implemented.

This covers about most of what we need to do various experiments. For with beam experiments we just need to to modify steps 1) and 2) and select input_nc to 8 in the train scripts. For selecting manual vs ECHO data for training we can change the "dataroot" for correct data (I have shared both ECHO/Manual training data). For selecting the loss function we need to modify step 3).