s1 - sort s2 - mask s3 - slice Input 3D Nii Keep slices with ROI's larger than 'threshold' value Create a square patch around the ROI (matrix size is a factor of 32) scale square patch to 64x64 Output 64x64 sized 2D NPY files
s4 - load Define how many dynamics we want (max = 3 dynamics) Load and normalize 2D npy arrays of patients that have desired number of dynamics Save all 2D npy arrays into a large 4D npy array (n,x,y,dynamics)
s5 - CNN Run CNN with data augmentation and 5 fold cross validation
# Block 1
Conv2D(64 filters, 3x3 kernel, activation='relu', padding='same', name='block1_conv2')
Conv2D(64 filters, 3x3 kernel, activation='relu', padding='same', name='block1_conv2')
MaxPooling2D(pool_size=(2, 2)
# Block 2
Conv2D(128 filters, 3x3 kernel, activation='relu', padding='same', name='block2_conv1')
Conv2D(128 filters, 3x3 kernel, activation='relu', padding='same', name='block2_conv2')
MaxPooling2D(pool_size=(2, 2))
# Block 3
Conv2D(256 filters, 3x3 kernel, activation='relu', padding='same', name='block3_conv1')
Conv2D(256 filters, 3x3 kernel, activation='relu', padding='same', name='block3_conv2')
Conv2D(256 filters, 3x3 kernel, activation='relu', padding='same', name='block3_conv3')
MaxPooling2D(pool_size=(2, 2)))
# Block 4
Conv2D(512 filters, 3x3 kernel, activation='relu', padding='same', name='block4_conv1')
Conv2D(512 filters, 3x3 kernel, activation='relu', padding='same', name='block4_conv2')
Conv2D(512 filters, 3x3 kernel, activation='relu', padding='same', name='block4_conv3')
MaxPooling2D(pool_size=(2, 2)))
Flatten()
Dense(4096, activation='relu')
Dropout(0.5)
Dense(num_classes, activation='softmax')
s6 - Create Accuracy and Loss Figure s7 - Generate Predictions s8 - Generate ROC Curve
Visualize 0 layer feature maps
