ircad data

Advisory

2022-08-01: Historical document included below. Links and directories may or may not work. No effort is undertaken to restore broken links, as this is document is now only for context.

Goal

Create a 3D solid model of the human head chest from medical images without human intervention.

Data

The 3D-IRCADb-01 database contains 3D CT scans of 10 males and 10 females with hepatic tumors in 75 percent of the cases. Patient-specific data follow, which is an extraction from the source table.

2022-08-01: Currently only patient_1 and patient_2 have been uploaded to the G Drive to save hard drive space. We are not certain if we will make use of the ircad data set because it is torso data. We want to focus on head data for application to brain injury, emphasized by ONR.

Patient No.	Sex	DOB	Voxel Size (mm)	Image Size (pixels)	G drive link
1	F	1944	0.57, 0.57, 1.6	512, 512, 129	patient_1
2	F	1987	0.78, 0.78, 1.6	512, 512, 172	patient_2
3	M	1956	0.62, 0.62, 1.25	512, 512, 200	to come
4	M	1942	0.74, 0.74, 2.0	512, 512, 91	to come
5	M	1957	0.78, 0.78, 1.6	512, 512, 139	to come
6	M	1929	0.78, 0.78, 1.6	512, 512, 135	to come
7	M	1946	0.78, 0.78, 1.6	512, 512, 151	to come
8	F	1970	0.56, 0.56, 1.6	512, 512, 124	to come
9	M	1949	0.87, 0.87, 2.0	512, 512, 111	to come
10	F	1953	0.73, 0.73, 1.6	512, 512, 122	to come
11	M	1966	0.72, 0.72, 1.6	512, 512, 132	to come
12	F	1973	0.68, 0.68, 1.0	512, 512, 260	to come
13	M	1951	0.67, 0.67, 1.6	512, 512, 122	to come
14	F	1970	0.72, 0.72, 1.6	512, 512, 113	to come
15	F	1946	0.78, 0.78, 1.6	512, 512, 125	to come
16	M	1950	0.70, 0.70, 1.6	512, 512, 155	to come
17	M	1942	0.74, 0.74, 1.6	512, 512, 119	to come
18	F	1958	0.74, 0.74, 2.5	512, 512, 74	to come
19	F	1970	0.70, 0.70, 4	512, 512, 124	to come
20	F	1949	0.81, 0.81, 2	512, 512, 225	to come

Potential New Data Source

CSI 2014

HPC SYN hx

Accessing the p2m directory

Our directory is located at data/wg-p2m. You can get to it from your home directory by running cd ../../data/wg-p2m.

Using the same python environment

From data/wg-p2m, run source miniconda/bin/activate followed by conda activate p2m.

IMPORTANT: Don't mess with the version of tensorflow, keras, or cuda. That took a while to set up, and while I [Anirudh] do have notes, recreating it might be a bit of a pain.

Data and Code

• The data are located in geometry/doc/pixel2mesh/unet/data/
• The code is located in geometry/doc/pixel2mesh/unet/unet_keras/

Code description

data_gen_np.py

This module defines the PatientData object.

• PatientData looks for files in the 3DIRCAD data folder. The database contains DICOM files located in different folders, PATIENT_DICOM which contains the original CT scans, and MASKS_DICOM, which contains subfolders of masks of the CT scan. For example, the directory MASKS_DICOM/bone will contain the corresponding bone masks for images in PATIENT_DICOM
• You can then call PatientData.get_data, which resizes and normalises the data before splitting it into train and test and saving them out as .npy files

make_data_np.py

This module creates a PatientData object and calls get_data to save the scans and masks as numpy files.

model.py

This module defines the function get_unet which returns a compiled U-Net model. A description of the model can be found in the file itself or in the U-Net paper.

losses.py

This module defines three functions.

1. dice_coef, which is defined as DSC = 2|X ∩ Y| / (|X| + |Y|)
2. dice_coef_loss, which defines the loss as 1-dice_coef. -dice_coef is equally acceptable, but less readable as we generally expect the loss to have converged at 0 and not -1
3. dice_coef_no_tf again defines the dice coefficient, but this time without using tensorflow functions. This is useful for visualizing results.

callbacks.py

This module implements the callbacks used during the training of our U-Net model. We use the standard ModelCheckpoint, ReduceLROnPlateau, and EarlyStopping callbacks, while defining an additional StopTraining callback, which stops the training if the dice coefficient reaches 1.0.

train_np.py

This module loads the numpy files and trains.

eval_model.py

This module evaluates our model on the test data.

make_predictions.py

This module predicts what the mask for a given scan will look like and saves the result to a .npy file

visualize_pred.py

This module shows a plot comparing our prediction with the true result.

patient_film.py

This module plots a nice animation scrolling through a patient's scans.