RSNA-MICCAI Brain Tumor Radiogenomic Classification Competition

Task: Predict the presence or absence of the MGMT promoter gene from the MRI images

Content

1. Preprocessed the train images of the competition using Torchio such that all have Axial view.
2. Then combined the (chosen slices) of FLAIR, T2w, and T1wCE into an image with 3 channels.
3. The size is 224 x 224 so that EfficientNetB0 can be used with input size (224, 224, 3).
4. Simple thresholding, median filtering, and morphological operations were used on the FLAIR MRI for the tumor segmentation. Please note that the tumor segmentations/localizations are not so clean. It is still recommended to only use the png images
5. The tumor segmentation masks are 2D numpy arrays.

Kaggle Notebook here

Acknowledgements

Thanks to Torchio and to this Notebook by Fernando Pérez-García Thanks everyone! 👍

Updates

• Version 2 now included higher resolution images (600 x 600) 🙌
• Version 3 considered the best slices (instead of the middlemost images) according to average pixel value of a FLAIR slice. The observation is that the FLAIR slice with the highest average pixel value contains the best tumor (edema) view.
• Version 4 improved the alignment of slices by stacking the slices before cropping.
• Version 5 Included test images, removed 600x600 images, included id=11, fixed the error in version_4 which deletes id=442

Sample Usage

> import numpy as np > import cv2 > case_id = '00000'

> image = cv2.imread(f'../input/brats21-preprocessed-train-images-with-torchio/preprocessed_images/{case_id}.png')

> tumor_mask = np.load(f'../input/brats21-preprocessed-train-images-with-torchio/tumor_mask/{case_id}.npy')