MRI_Segmentation_Radiomic / Git / Diff of /README.md

Models:

SCallahan/

MRI_Segmentation_Radiomic

Downloads: 1

Diff of /README.md [7c5f70] .. [d14a9a]

Switch to unified view


![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/CIVMBanner.png)
# MRI Segmentation and Radiomics
This repository contains example code from the paper in preparation on preclinical cancer imaging titled "MRI-based 
Deep Learning Segmentation and Radiomics of Sarcoma Tumors in Mice."

This work is part of the [U24 co-clinical trial](https://sites.duke.edu/pcqiba/) of which [CIVM](http://www.civm.duhs.duke.edu/) at Duke University is a participant. This work has been funded by **NIH U24CA220245**.


This goal of this project is to create an pipeline with two facets:
1. Automatic segmentation of soft tissue sarcoma using deep learning.
2. Automate calculation and analysis of radiomic features within the tumor population.

## Data
The data used in this study is composed of multi-contrast MR images of soft tissue sarcoma. Tumors were imaged using 
T2-weighted and T1-weighted sequences. These were followed by a contrast-enhanced T1-weighted acquisition.
![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/multi_contrast.png)

The full dataset will soon be available on the [CIVM VoxPort page](https://civmvoxport.vm.duke.edu/voxbase/studyhome.php?studyid=617)

## Segmentation
Segmentation was performed via a U-net CNN. The network functions on patches taken from image volumes. The general 
network structure is shown below.
![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/cnn_structure.png)

Training and perfomance anlysis is done using the [Segmentation.py](Segmentation/Segmentation.py) script. The results
 for a network trained on multi-contrast MR images with cross entropy loss is shown below.
 
 ![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/segmentations.png)

#### Requirements
T2-weighted images are bias corrected using N4BiasFieldCorrection in [ANTs](http://stnava.github.io/ANTs/).

Network structures have been defined in [model_keras.py](Segmentation/model_keras.py). A variety of networks have 
been tested, including those with and without skip connections as well as testing out cross entropy and dice loss 
functions.



## Radiomics
Radiomic features were calculated within the segmentation using [PyRadiomics](https://pyradiomics.readthedocs.io/en/latest/). 
Radiomic 
calculation are performed using [radiomic_calculations_batch.py](Radiomics/radiomic_calculations_batch.py).

The code used to process the radiomic features is found in [radiomic_classifications.py]
(Radiomics/radiomics_classifications.py). The features are plotted and classified using SVMs 
and NNs to determine if primary local recurrence can be predicted in this population based only on radiomic features.
 Machine learning classifier functions lie in [radiomic_functions.py](Radiomics/radiomic_functions.py).
 
Using this code, we achieved an AUC of 0.81 for predicting recurrence within these mice.
 
![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/classifier_results.png)
 
#### Requirements
Due to the high dimensionality of the data (321 features per tumor) feature selection is required
. This are accomplished via [mRMR](http://home.penglab.com/proj/mRMR/), which must be added to the path before running. 

## Crawler
A crawler has been implemented to locate, pre-process, segmentent, and compute radiomic features for large 
collections of images. The code for this located in the "Crawler" folder and can be run through the [crawler.py](Crawler/crawler.py)
script once the base paths have been updated.

	a/README.md		b/README.md
1	![image](.github/CIVMBanner.png)	1	![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/CIVMBanner.png)
2	# MRI Segmentation and Radiomics	2	# MRI Segmentation and Radiomics
3	This repository contains example code from the paper in preparation on preclinical cancer imaging titled "MRI-based	3	This repository contains example code from the paper in preparation on preclinical cancer imaging titled "MRI-based
4	Deep Learning Segmentation and Radiomics of Sarcoma Tumors in Mice."	4	Deep Learning Segmentation and Radiomics of Sarcoma Tumors in Mice."
5		5
6	This work is part of the [U24 co-clinical trial](https://sites.duke.edu/pcqiba/) of which [CIVM](http://www.civm.duhs.duke.edu/) at Duke University is a participant. This work has been funded by NIH U24CA220245.	6	This work is part of the [U24 co-clinical trial](https://sites.duke.edu/pcqiba/) of which [CIVM](http://www.civm.duhs.duke.edu/) at Duke University is a participant. This work has been funded by NIH U24CA220245.
7		7
8		8
9	This goal of this project is to create an pipeline with two facets:	9	This goal of this project is to create an pipeline with two facets:
10	1. Automatic segmentation of soft tissue sarcoma using deep learning.	10	1. Automatic segmentation of soft tissue sarcoma using deep learning.
11	2. Automate calculation and analysis of radiomic features within the tumor population.	11	2. Automate calculation and analysis of radiomic features within the tumor population.
12		12
13	## Data	13	## Data
14	The data used in this study is composed of multi-contrast MR images of soft tissue sarcoma. Tumors were imaged using	14	The data used in this study is composed of multi-contrast MR images of soft tissue sarcoma. Tumors were imaged using
15	T2-weighted and T1-weighted sequences. These were followed by a contrast-enhanced T1-weighted acquisition.	15	T2-weighted and T1-weighted sequences. These were followed by a contrast-enhanced T1-weighted acquisition.
16	![image](.github/multi_contrast.png)	16	![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/multi_contrast.png)
17		17
18	The full dataset will soon be available on the [CIVM VoxPort page](https://civmvoxport.vm.duke.edu/voxbase/studyhome.php?studyid=617)	18	The full dataset will soon be available on the [CIVM VoxPort page](https://civmvoxport.vm.duke.edu/voxbase/studyhome.php?studyid=617)
19		19
20	## Segmentation	20	## Segmentation
21	Segmentation was performed via a U-net CNN. The network functions on patches taken from image volumes. The general	21	Segmentation was performed via a U-net CNN. The network functions on patches taken from image volumes. The general
22	network structure is shown below.	22	network structure is shown below.
23	![image](.github/cnn_structure.png)	23	![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/cnn_structure.png)
24		24
25	Training and perfomance anlysis is done using the [Segmentation.py](Segmentation/Segmentation.py) script. The results	25	Training and perfomance anlysis is done using the [Segmentation.py](Segmentation/Segmentation.py) script. The results
26	for a network trained on multi-contrast MR images with cross entropy loss is shown below.	26	for a network trained on multi-contrast MR images with cross entropy loss is shown below.
27		27
28	![image](.github/segmentations.png)	28	![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/segmentations.png)
29		29
30	#### Requirements	30	#### Requirements
31	T2-weighted images are bias corrected using N4BiasFieldCorrection in [ANTs](http://stnava.github.io/ANTs/).	31	T2-weighted images are bias corrected using N4BiasFieldCorrection in [ANTs](http://stnava.github.io/ANTs/).
32		32
33	Network structures have been defined in [model_keras.py](Segmentation/model_keras.py). A variety of networks have	33	Network structures have been defined in [model_keras.py](Segmentation/model_keras.py). A variety of networks have
34	been tested, including those with and without skip connections as well as testing out cross entropy and dice loss	34	been tested, including those with and without skip connections as well as testing out cross entropy and dice loss
35	functions.	35	functions.
36		36
37		37
38		38
39	## Radiomics	39	## Radiomics
40	Radiomic features were calculated within the segmentation using [PyRadiomics](https://pyradiomics.readthedocs.io/en/latest/).	40	Radiomic features were calculated within the segmentation using [PyRadiomics](https://pyradiomics.readthedocs.io/en/latest/).
41	Radiomic	41	Radiomic
42	calculation are performed using [radiomic_calculations_batch.py](Radiomics/radiomic_calculations_batch.py).	42	calculation are performed using [radiomic_calculations_batch.py](Radiomics/radiomic_calculations_batch.py).
43		43
44	The code used to process the radiomic features is found in [radiomic_classifications.py]	44	The code used to process the radiomic features is found in [radiomic_classifications.py]
45	(Radiomics/radiomics_classifications.py). The features are plotted and classified using SVMs	45	(Radiomics/radiomics_classifications.py). The features are plotted and classified using SVMs
46	and NNs to determine if primary local recurrence can be predicted in this population based only on radiomic features.	46	and NNs to determine if primary local recurrence can be predicted in this population based only on radiomic features.
47	Machine learning classifier functions lie in [radiomic_functions.py](Radiomics/radiomic_functions.py).	47	Machine learning classifier functions lie in [radiomic_functions.py](Radiomics/radiomic_functions.py).
48		48
49	Using this code, we achieved an AUC of 0.81 for predicting recurrence within these mice.	49	Using this code, we achieved an AUC of 0.81 for predicting recurrence within these mice.
50		50
51	![image](.github/classifier_results.png)	51	![image](https://github.com/mdholbrook/MRI_Segmentation_Radiomics/blob/master/.github/classifier_results.png)
52		52
53	#### Requirements	53	#### Requirements
54	Due to the high dimensionality of the data (321 features per tumor) feature selection is required	54	Due to the high dimensionality of the data (321 features per tumor) feature selection is required
55	. This are accomplished via [mRMR](http://home.penglab.com/proj/mRMR/), which must be added to the path before running.	55	. This are accomplished via [mRMR](http://home.penglab.com/proj/mRMR/), which must be added to the path before running.
56		56
57	## Crawler	57	## Crawler
58	A crawler has been implemented to locate, pre-process, segmentent, and compute radiomic features for large	58	A crawler has been implemented to locate, pre-process, segmentent, and compute radiomic features for large
59	collections of images. The code for this located in the "Crawler" folder and can be run through the [crawler.py](Crawler/crawler.py)	59	collections of images. The code for this located in the "Crawler" folder and can be run through the [crawler.py](Crawler/crawler.py)
60	script once the base paths have been updated.	60	script once the base paths have been updated.