Technical documentation: detect_cmb

detect_cmb

dataloader

###savepatches.py

saves ground truth and corresponding input patches in a separate directory

---

###hdf5_convert.py

converts the input and ground truth patches to a single hdf5 file containing subgroups of 'input_images,'output_images'

---

###main_hdf5.py

main method for accessing hdf5 file and executing screening stage 1 and 2

---

balanced_dataset.py

def balanced_dataset(subjects):

subjects = 3D patches of input and ground truth images

• Prepare dataset with an equal number of positive and negative patches

---

create_dataloader.py

def trainTestSplit(self, dataset, val_share):

dataset = complete data set after preprocessing

val_share = validation data set size

• Return training and validation set

---

def create_dset_screening_stage1(self,dset):

• Convert the data set into torch format

---

dataloader.py

Split the complete data set into train, validation and test sets based on number of CMB patches in each subject

def get_split(self,input_dir):

input_dir = file location of home directory

• Read the ground truth of each subject
• If the ground truth has a CMB, append it to a dictionary (subject, number of CMB in that subject)
• Now split the dictionary based on a fixed percentage of train, validation and test set ratio

---

discrimination_dataset.py

def prepare_datset_with_mimics(subjects, checkpoint):

subjects = list of train subjects checkpoint = trained model from screening stage-2

• Evaluate the train subjects based on the trained model of screening stage-2
• Find out positives and false positives based on this evaluation
• This dataset is used for training the discrimination stage

---

generate_patches.py

def cut_3d_image(self, index_x, index_y, index_z, size_x, size_y, size_z, x_stride, y_stride, z_stride, img, image,gt):

• Cut the 3D volumetric data into 3D patches of 16 x 16 x 10 size
• Return the patches - these patches are used for training

---

def create_3dpatches(self,subjects):

subjects = preprocessed x,y pair where x = input and y = ground truth

• Define the size of the 3D patch here
• Invoke cut_3d_image() to create 3D patches

---

executor

screening_stage.py

• train and validate the screening stage-1 and checkpoints for each epoch is saved

screening_stage_2.py

• train and validate the screening stage-2 and checkpoints for each epoch is saved

model

screening.py

• defines network for screening stage

discrimination.py

• defines network for discrimination stage

ops

dicom_to_nifti.py

• Convert the DICOM to NIfTI images

extract_mimics.py

• Find false positives based on a model trained from screening stage-1

generateGroundTruth.py

• Create ground truth based on CMB co-ordinates in meta-file

normalize_images.py

• Normalize the intensity of input NIfTI images

score_map.py

• To find CMB candidate based on a model trained from screening stage-2