real-time OCT denoising implemented on SECTR imaging system

Under /utils/ there are some functions used. util.py includes the Nifiti file reader/writer, intensity normalizer etc. MotionCorrection.py is the function used for rigid registration. label_fusion is joint label fusion sofware. self_fusion.sh is a shell script that inplement the label fusion software. There are 3 directories to be specified, dir of label_fusion, dir of c3d, dir of atlases (bscans in our case).

The model Nsf is used to map the high noise input to the self-fusion of the corresponding low noise image . In /src/Nsf_train/, all the pre-processing and model training code is included.

(1) train_data_prep.py

This file is used to get n-frame average by (a) shown in the pipleline. Also, the 1st frame of the n repeats is extracted. Other than this, it register r-neighboring bscans of and save it as a stack of (2r+1) channels for later self-fusion. In conclusion, there are 3 outputs for each raw volume, X,Y (Nifiti) and a pickle file that contains the low noise (Y) bscans stacks.

      X: HN_ONH_SNR_101_1.nii.gz  
      Y: LN_ONH_SNR_101_1.nii.gz
      Y Stack: ONH_SNR_101_1.pickle

(2) Get_SFy.py

Given the aligned r-neighborhood of , the self-fusion is applied to get . We call the subprocess to execute self_fusion.sh. Note that the directory of atlases should be consistant with $Dir in self_fusion.sh.

(3) PM_dataloader.py

The PM (stands for pseudo-modality) dataloader is used to pair (,). In order to provide sufficient information for the model to mimic the self-fusion. The input has 3 channels including ,,. Adjacent bscans are also registered. The output will be a pickled list. Each element of the list will be a (X,S) pair.

(4) PM_train.py

The training of the model. The architecture of the model is the multi-scale U-Net which is available under /Models/NetworkArch.py

To train Ndn, the input X has 3 channels and supervised by both and .

    X: channel 1 = high noise bscan 
       channel 2 = output of Nsf
       channel 3 = sobel edge
    Y: y1 = n-frame average
       y2 = self-fusion of n-frame average

(1) Ndn_dataloader.py

This file is used to get Nsf(x) and sobel edge, stack them with the high noise bscan and pair with ground truth.

(2) Ndn_train.py

This is the training code for Ndn. The network architecture is also multi-scale U-Net (MSUN).

The full testing of the whole pipeline is done by main.py. Some parts are hard coded, it is now fit nch1=3 for Nsf (using 3 bscans to mimic self-fusion with 7 bscans). nch2=3 because the input of Ndn include (high noise bscan, SF, sobel edge).

(1) The network architechture

(2) trained model