Deep Learning Based Segmentation of Multicontrast TNBC PDX MR Images

This repository provides an automated pipeline for automatic localization and segmentation of multicontrast MR images in small animals. In this project we have tried using UNet ¹, Residual UNet ², Dense UNet ³, Recurrent Residual UNet (R2UNet) ⁴ and Dense Recurrent Residual UNet (DR2UNet) ⁵.

Steps to Test the Deep Learning Pipeline for Segmentation

1. Git Clone the repository
2. Install python 3.7.10 and the necessary packages by running pip install -r requirements.txt
3. Download the trained weights from https://wustl.box.com/s/hntgs87p91fscx624vpajru6xa789p6m into a weights_new folder inside the folder where you have cloned the other codes from the repository.
4. Download images_test.npy , images_train.npy, images_train_T1.npy and mask_train.npy.
5. Run main_testing.py. Have to manually load which model you want to learn. By default the DR2UNet is loaded and running the main_testing.py gives the performance score for the selected network on the test dataset. It automatically calculates the optimum threshold based on the precision recall curves for the network. The main_testing.py automatically calls the model architecture and the data2D.py to load the data.
6. You can view the output probability maps in output folder under each model folder name (which you have to change each time you run a model).

Different Network Model and their architecture files:

Network Model	Architecture File	Weight File
DR2UNet	new_r2udensenet.py	model_r2udensenet.hdf5
R2UNet	new_r2unet.py	model_r2unet.hdf5
Dense-UNet	dense_unet.py	model_denseunet.hdf5
Res-UNet	res_unet.py	model_resunet.hdf5
UNet	unet.py	model_unet.hdf5

Steps for Training and Five Fold Cross Validation

1. Git Clone the repository.
2. Install python 3.7.10 and the necessary packages by running pip install -r requirements.txt
3. For five fold cross validation run main2D.py after selecting the network model you want to train the data on. The training considers data augmentation due to the limited dataspace available to us. The main2D.py outputs the mean performance of the network model for 5 folds.
4. For standalone training on the whole dataset after the cross-validation run main_training.py to generate the training-weights.

References

1. Ronneberger, O.; Fischer, P.; Brox, T., U-net: Convolutional networks for biomedical image segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 2015, 9351, 234-241.
2. He, K.; Zhang, X.; Ren, S.; Sun, J., Deep residual learning for image recognition. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2016, 2016-Decem, 770-778.
3. Kolařík, M.; Burget, R.; Uher, V.; Říha, K.; Dutta, M. K., Optimized high resolution 3D dense-U-Net network for brain and spine segmentation. Applied Sciences (Switzerland) 2019, 9 (3).
4. Alom, M. Z.; Hasan, M.; Yakopcic, C.; Taha, T. M.; Asari, V. K., Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) for Medical Image Segmentation. 2018.
5. Dutta, K., Densely Connected Recurrent Residual (Dense R2UNet) Convolutional Neural Network for Segmentation of Lung CT Images. arXiv preprint arXiv:2102.00663 2021.