BME-570-project-code

1. Matlab code to evaluate performance on a signal-detection task:

   ./observer_evaluation/performance_evaluation.m

2. Python code to build a CNN for denoising:

   ./denoising/DL_denoiser.py

Observer Evaluation

Usage

  [AUC] = performance_evaluation(IS,IN)

Discription

Conduct an observer study using channelized Hotelling observer. The signal-present images are given by IS, the signal-absent are images given by IN. Then compute the AUC, and plot the ROC curves (by simple thresholding).

Inputs

  IS -- [Number_of_Pixels X Number_of_Images] -- The input signal-present images 
            (Each column represents one image)
  IN -- [Number_of_Pixels X Number_of_Images] -- The input signal-absent images 
            (Each column represents one image)

Outputs

  AUC -- Wilcoxon area under the ROC curve
  Figures -- ROC curve and delta f_hat_bar image.
  (delta f_hat_bar: the difference between signal-present and signal-absent images)

Example of output

Matlab command line output:

  AUC = 0.9530

More to read

Harrison H. Barrett and Kyle J. Myers, Foundations of Image Science, Charpter 13.2

---

Edited by Zitong Yu @ Nov-16-2021

Based on IQmodelo toolbox Apr-15-2016© the Food and Drug Administration (FDA) and IMAGE QUALITY TOOLBOX Version 0.9b Mar-25-2001© The University of Arizona

Contact: [email protected]

CNN-based denoiser

A simple CNN-based denoiser. Feel free to change the structure of the network. If you find any bug or have any question, please contact Zitong Yu ([email protected]). Thanks.

Python requirements

This code was tested on:

• Python 3.6
• TensorFlow 1.10.0
• Keras 2.2.4

No GPU requirement, since the code runs on CPUs.

Preparing dataset

Images should be saved as 2-D .npy files. Use provided writeNPY() MATLAB function to save images as .npy files:

  writeNPY(reshape(img,32,32),['image',num2str(img_ind),'.npy']);

The training data should be saved as following:

$training_data_dir$ -- noisy
                        -- image0.npy
                        -- image1.npy
                        -- ...
                    -- low_noise
                        -- image0.npy
                        -- image1.npy
                        -- ...

The test data should be saved as follwoing:

$test_data_dir$ -- noisy
                  -- image0.npy
                  -- image1.npy
                  -- ...

Running

Run python DL_denoiser.py -h for a complete listing of command line parameters and support list of training configurations.

usage: DL_denoiser.py [-h] [--dataset-dir DATASET_DIR]
                      [--trained-model TRAINED_MODEL] [--num-train NUM_TRAIN]
                      [--num-test NUM_TEST] [--num-epochs NUM_EPOCHS]
                      [--dim DIM] [--batch-size BATCH_SIZE]

Train the CNN or do the perdiction.

optional arguments:
  -h, --help            show this help message and exit
  --dataset-dir DATASET_DIR
                        Path to dataset.
  --trained-model TRAINED_MODEL
                        Trained model.
  --num-train NUM_TRAIN
                        Number of traing data.
  --num-test NUM_TEST   Number of testing data.
  --num-epochs NUM_EPOCHS
                        Number of epochs.
  --dim DIM             Image size. For example, if (32,32) image, dim = 32.
  --batch-size BATCH_SIZE
                        Batch size.

Training

To train a network, run:

python DL_denoiser.py --dataset-dir $your_dataset_directory$ --num-train $number_of_training_samples$ --num-epochs $number_of_epochs$ --dim $image_size$ --batch-size $batch_size$

An example:

python DL_denoiser.py --dataset-dir ./data/train/ --num-train 400 --num-epochs 100 --dim 32 --batch-size 8

A successful training should print the following (64X64 images):

Using TensorFlow backend.
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to
==================================================================================================
input_1 (InputLayer)            (None, 64, 64, 1)    0
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 64, 64, 16)   160         input_1[0][0]
__________________________________________________________________________________________________
leaky_re_lu_1 (LeakyReLU)       (None, 64, 64, 16)   0           conv2d_1[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout)             (None, 64, 64, 16)   0           leaky_re_lu_1[0][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 32, 32, 16)   1040        dropout_1[0][0]
__________________________________________________________________________________________________
leaky_re_lu_2 (LeakyReLU)       (None, 32, 32, 16)   0           conv2d_2[0][0]
__________________________________________________________________________________________________
dropout_2 (Dropout)             (None, 32, 32, 16)   0           leaky_re_lu_2[0][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 32, 32, 32)   4640        dropout_2[0][0]
__________________________________________________________________________________________________
leaky_re_lu_3 (LeakyReLU)       (None, 32, 32, 32)   0           conv2d_3[0][0]
__________________________________________________________________________________________________
dropout_3 (Dropout)             (None, 32, 32, 32)   0           leaky_re_lu_3[0][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D)               (None, 16, 16, 32)   4128        dropout_3[0][0]
__________________________________________________________________________________________________
leaky_re_lu_4 (LeakyReLU)       (None, 16, 16, 32)   0           conv2d_4[0][0]
__________________________________________________________________________________________________
dropout_4 (Dropout)             (None, 16, 16, 32)   0           leaky_re_lu_4[0][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D)               (None, 16, 16, 64)   18496       dropout_4[0][0]
__________________________________________________________________________________________________
leaky_re_lu_5 (LeakyReLU)       (None, 16, 16, 64)   0           conv2d_5[0][0]
__________________________________________________________________________________________________
dropout_5 (Dropout)             (None, 16, 16, 64)   0           leaky_re_lu_5[0][0]
__________________________________________________________________________________________________
conv2d_transpose_1 (Conv2DTrans (None, 32, 32, 32)   8224        dropout_5[0][0]
__________________________________________________________________________________________________
leaky_re_lu_6 (LeakyReLU)       (None, 32, 32, 32)   0           conv2d_transpose_1[0][0]
__________________________________________________________________________________________________
add_1 (Add)                     (None, 32, 32, 32)   0           leaky_re_lu_6[0][0]
                                                                 dropout_3[0][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D)               (None, 32, 32, 32)   9248        add_1[0][0]
__________________________________________________________________________________________________
leaky_re_lu_7 (LeakyReLU)       (None, 32, 32, 32)   0           conv2d_6[0][0]
__________________________________________________________________________________________________
dropout_6 (Dropout)             (None, 32, 32, 32)   0           leaky_re_lu_7[0][0]
__________________________________________________________________________________________________
conv2d_transpose_2 (Conv2DTrans (None, 64, 64, 16)   2064        dropout_6[0][0]
__________________________________________________________________________________________________
leaky_re_lu_8 (LeakyReLU)       (None, 64, 64, 16)   0           conv2d_transpose_2[0][0]
__________________________________________________________________________________________________
add_2 (Add)                     (None, 64, 64, 16)   0           leaky_re_lu_8[0][0]
                                                                 dropout_1[0][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D)               (None, 64, 64, 16)   2320        add_2[0][0]
__________________________________________________________________________________________________
leaky_re_lu_9 (LeakyReLU)       (None, 64, 64, 16)   0           conv2d_7[0][0]
__________________________________________________________________________________________________
dropout_7 (Dropout)             (None, 64, 64, 16)   0           leaky_re_lu_9[0][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D)               (None, 64, 64, 1)    145         dropout_7[0][0]
__________________________________________________________________________________________________
leaky_re_lu_10 (LeakyReLU)      (None, 64, 64, 1)    0           conv2d_8[0][0]
==================================================================================================
Total params: 50,465
Trainable params: 50,465
Non-trainable params: 0
__________________________________________________________________________________________________
Loading the training data...
Training...
Train on 300 samples, validate on 100 samples
Epoch 1/100
38/38 [==============================] - 3s 78ms/step - loss: 21.6744 - val_loss: 20.1661

<... snip ...>

Training finished.

A trained model will be saved in the current directory.

Testing

To test a network, run:

python DL_denoiser.py --dataset-dir $your_dataset_directory$ --num-test $number_of_test_samples$ --dim $image_size$ --batch-size $batch_size$ --trained-model $the_trained_model$

An example:

python DL_denoiser.py --dataset-dir ./data/test/ --num-test 50 --dim 32 --batch-size 8 --trained-model ./model_epochs100.h5

A successful training should print the following:

Using TensorFlow backend.

<... Model structure, snip ...>

Total params: 50,465
Trainable params: 50,465
Non-trainable params: 0
__________________________________________________________________________________________________
Loading the test data...
100/100 [==============================] - 1s 13ms/step
Saving the predicted data...

The predictions will be saved in the $your_dataset_directory$/prediction.

Use provided readNPY() MATLAB function to read images in MATLAB:

  img = readNPY(['image',num2str(img_ind),'.npy']);

---

Edited by Zitong Yu @ Nov-19-2021

Contact: [email protected]