GBM_WSSM

• Model Name: GBM_WSSM
• "Glioblastoma Whole-Slide Image Semantic Segmentation Model using Deep Convolutional Neural Network"

-----------------------------------------------

-----------------------------------------------

Dependencies:

• Python: 3.6
• Tensorflow: 1.13.1
• Keras: 2.2.4

-----------------------------------------------

Steps to reproduce the results of the paper

1. Download the optimum weights of "GBM_WSSM" trained on IVY-GAP dataset from the link "GBM_WSSM.h5" and place it in its appropriate part in the code "GBM_WSSM_Prediction.py".

Note: Color codes for eight different regions (IT, LE, CT, CTmvp, CTne, CTpnz, CTpan, and Background) are located in the Mask_Labels.txt.

2. Download GBM test images from this link or use your own GBM images. For pre-processing of your images, please refer to the original article. Then, run the code "GBM_WSSM_Prediction.py". you can compare the results (produced masks by GBM_WSSM) with the original ground truth in this link.

Arguments:

GBM_Parser.add_argument("--Image_Info",
                        help = "This part includes some initial information about the image you want to create mask for: shape, number of classes",
                        default = [(1024, 1024, 3), 8])
GBM_Parser.add_argument("--GBM_Test_Images_Path",
                        help = "The path of 48 GBM images in testing phase",
                        default = ".../.../Test/Images/")

GBM_Parser.add_argument("--Predicted_Masks_Path",
                        help = "The path of the Predicted Masks by GBM_WSSM",
                        default = ".../.../")

GBM_Parser.add_argument("--GBM_WSSM_h5_path",
                        help = "GBM_WSSM best model weights.h5 path",
                        default = ".../.../GBM_WSSM.h5")

GBM_Parser.add_argument("--Labels_txt",
                        help = "Mask_Labels.txt path",
                        default = ".../.../Mask_Labels.txt")

-----------------------------------------------

Train the DCNN model from scratch

1. If you want to train the model from scratch, download raw images and corressponding masks from the link . Also, you can save your time and train the model by using training and validating numpy arrays we prepared. For both ways, first download the images or arrays from and put them in the main_code.

Arguments:

GBM_Parser.add_argument("--GBM_Images_Path",
                        help = "The path of the original GBM images",
                        default = ".../.../")

GBM_Parser.add_argument("--GBM_GTs_Path",
                        help = "The path of the original GBM Ground Truth (GTs)",
                        default = ".../.../")

GBM_Parser.add_argument("--GBM_Regions_and_RGBs_List",
                        help = "A text file which defines all the regions (classes) inside GBM GTs along with their RGB vlaues i.e. Mask_Labels.txt",
                        default = ".../.../")

GBM_Parser.add_argument("--Encode_GBM_GTs",
                        help = "The default value is False. If you change it to True, all the GTs images will be mapped (RGBs will be encoded to integers) using Mask_Integer_Encoding()",
                        default = True)

GBM_Parser.add_argument("--Model_Outputs_Path",
                        help = "In this path, all the model outputs will be saved (including: data arrays, curves, weights.hdf5, ...",
                        default = ".../.../")

GBM_Parser.add_argument("--Log_Files_Path",
                        help = "This path will save callbacks outputs: checkpoint files (*.h5 or *.hdf5 formats), TensorBoard records, and events from CSVLogger (*.csv)",
                        default = ".../.../")

GBM_Parser.add_argument("--Hyper_parameters",
                        help = "A list that contains all necessary hyperparameters: Learning_Rate, Epochs, Patience, LR_Decay, Train_Percentage, Batch_Size, and class no., respectively",
                        default = [1e-3, 55, 50, 5e-5, 85, 1, 8])

GBM_Parser.add_argument("--GBM_Image_Size",
                        help = "The new size of GBM images. The size can be used if you want to use Resize function. Default is (1024, 1024)",
                        default = (1024, 1024))

GBM_Parser.add_argument("--Image_Size_Boolean",
                        help = "Image_Size_Boolean is an argument whcih is being used whenever we need to resize the originl images or GTs",
                        default = True)

GBM_Parser.add_argument("--Cropped_Size",
                        help = "Crop_Size is used as a predefined size for crop extraction in Augmentaion procedure. Default is (224, 224)",
                        default = (224,224))

GBM_Parser.add_argument("--Resume",
                        help = "The default value is False, If you change it to True, the training process will be resumed provided you saved the weights", 
                        default = False)

GBM_Parser.add_argument("--Start_from_Scratch",
                        help = "If you are going to start from scratch i.e. make up training, and validating sets, and encode all masks, and train the model, apply --> TRUE. Otherwise, if you apply False, just predefined arrays are loaded from disk and the model will be trained on", 
                        default = False)

-----------------------------------------------

Steps to generalize GBM_WSSM on TCGA dataset

1. Original Images (.SVS format) for all types of brain tumours are accessible via the link.

2. We extracted only GBM tumour types using the GBM Log File and GBM Finder Code. GBM images (.SVS format) are accessible via the link.

3. All the GBM tumours were visually checked by a pathologist and whole slide images (WSIs) for each GBM are accessible via the link.

4. Resized form (1024x1024) of the GBM WSIs are available in this link.

5. Corresponding produced masks (semantically segmented images) by applying GBM_WSSM on GBM WSIs are accessible in the link.

6. This code has been used to quantify regions in the segmented masks obtained from the previous step. The quantified values are accessible here.

-----------------------------------------------

Notes

Deep Convolutional Neural Network Architecture for GBM Semantic Segmentation is inspired by the Article