The repository for the project in TDT4265 the spring of 2021 by group 3
This Project has been developed by:
Gulleik L Olsen [email protected] Software Engineer
Erling Olweus [email protected] Software Engineer
Fredrik Gran-Jansen [email protected] Software Engineer
In this project we will go through a typical medical imaging workflow often seen in real world
scenarios. The main task in the project will be segmentation of structures of interest in ultrasound
images of the heart (see figure). Two modalities will be considered, transthoracic ultrasound
(with ground truth segmentations available) and trans esophageal ultrasound (expert validated
test set is available).
We were free to implement our solution however we wanted, but there were some system requirements which had to be fulfilled:
-
DICE score higher than baseline code for TTE.
Implement a dice score (do it yourself or use one from a python package) for the sample code (that will be considered the baseline of 90%). Estimate a dice score for the left ventricle (LV), compare it to your work to it, describe what you did to improve the dice.
-
Model tested and working on TEE images.
Test your trained TTE model on the 19 TEE images with ground truth (you can discard the h5 files they were given as sample until the other ones were ready).
-
Modify the network to be able to segment all 3 classes
The classes are the left ventricle, myocardium and left atrium on the original CAMUS data. Recommended test set is patients 401-450.
-
Rotate the TEE data by 90 degrees in order to make it more similar to TTE data
Generate TTE and TEE images that are most similar content wise
-
Convert the data to iso-tropic pixel size
The element spacing should be equal in both directions.
-
Try and report Dice score on various input resolutions of your data
Subsample or oversample the images and the ground truth and then see what the network achieves in terms of dice score.
-
Use both the ED and ES timepoints from the TTE data for training purposes
This is regarding the CAMUS dataset where you need to use all 4 images per patient (4CH and 2CH both at ED and ES).
-
Report the effect of image smoothing on the Dice score (Gaussian or bilateral smoothing are 2 possible options)
Use some sort of smoothing filter on the image.
-
Surprise us with an idea you have
This is an idea that you have/want to try out that is not part of the requirements.
Clone this repostiory:
git clone https://github.com/gg1977/Unet2DYou can also download this repository as a zip file and unzip it on your computer.
This program was run in an anaconda environment with python 3.6.12. conda install --name myenv --file requirements.txt in order to install all required packages.
in order for the program to find the datafiles create a .env file with the paths to the data: e.g : TTE_BASE_PATH = "Data/CAMUS_resized/training" TTE_FULL_BASE_PATH = "Data/CAMUS/training" TTE_FULL_TEST_BASE_PATH = "Data/CAMUS/testing" TEE_BASE_PATH = "Data/DataTEEGroundTruth" The folder the path is pointing to should contain 1 folder for each patient containing all of the images. The testing folder contains medical images from patient 401-450. The TEE folder shold cotain two folders named train_gray and train_gt which contains the images and ground_truth respectivly.
In order to run the network with the default parameters run train.py configs/defaults.yaml
Loads data from TTE, TEE and CAMUS. Creates a dictionary with the paths to TTE and TEE images with corresponding ground truth image from the CAMUS dataset. Combines the ground truth and gray files. Reads the images with SimpleITK, creates an array containing the images and opens mask file.
Architecture with 4 down and 4 up convolutional layers. The contraction blocks and extraction blocks each have 2 convolutional layers, each followed by a BatchNorm2d and a ReLU function. The contraction blocks ends with a MaxPool2d function, while the extraction blocks ends with a ConvTranspose2d function.
We use albumentation.ai for augmentation, such as Rotate(), GaussianBlur() and Blur().
A collection of everything corresponding to training such as logging, outputs and checkpoints.
The logic behind training the model such as running through epochs. This module does all the traning.
Config file conatining the deafult configuration of our system (hyperparameters and so on)
File containing all the checkpointing functions, this file is copied from assigment 4 in https://github.com/hukkelas/TDT4265-StarterCode
File containing functions for dice calculations
File containing all the logging functions, this file is copied from assigment 4 in https://github.com/hukkelas/TDT4265-StarterCode
Suport functions for the checkpointer, this file is copied from assigment 4 in https://github.com/hukkelas/TDT4265-StarterCode
Different custom plotting functions
All our contraction and expansion blocks uses 3x3 kernel size and padding=1, except in our first contraction block, where we use kernel size 7x7 and padding=1.
