Gliomas are the most common primary brain malignancies, with different degrees of aggressiveness, variable prognosis and various heterogeneous histological sub-regions. Ample multi-institutional routine clinically-acquired multi-parametric MRI (mpMRI) scans of glioma, with pathologically confirmed diagnosis are used as the training, validation, and testing data for BraTS challenge 2021. Ground truth annotations of the tumor sub-regions are created and approved by expert neuroradiologists for every subject included in the training, validation, and testing data sets to quantitatively evaluate the predicted tumor segmentation’s. The dataset of BraTS 2020 contains data of 660 cases that is like 2640 mpMRI images.The present method for diagnosis of tumor is reactive, so once the patients show symptoms they get MRI and CT scans done. The scans are looked into by doctor and the tumor is confirmed but sometimes due to human error or tumor being small in size is not visible by naked eyes[Refer Fig. 1]. Thus in that case the role of Artificial Intelligence comes into picture. By using this approach the accuracy of prediction could be increased. This approach introduced the use of an ensemble classifier composed of different deep convolutional neural networks (CNN) architectures. The paper expects the accuracy to be higher in terms of prediction so as to get better accuracy so the tumor can be predicted in better way that is not even visible to naked eye.
The images from the data will be fed as input to the model, the images has to be prepossessed due to certain abnormalities like, presence of writing, excess black part or inversion of image, etc. As per decided the process done on image will be contouring, cropping, resizing, normalisation and on some images HoG (Histogram of oriented Gradient). The image converted to size of 240X240X1 matrix will be passed through CNN model consisting of 27 convolutional layers. The first half of the model will be used to narrow down the image to 15X15X256 and then the image will again be concatenated in 240X240X1 by using biases from the narrowing down order and by using total of 2,158,417 parameters to train the model. The MRI image are scanned that contains 5 types of images so if we take 20 images it scans a total of 100 images, these images are stored in for matrix, the image is normalized, the data is then processed, All the images are then concatenated into a single matrix and convolution is performed on it. Various other features like sensitivity, precision, specificity and dice coefficient are calculated based on the values; this represents the performance of the model.
Segmented Part of Tumour
Loss Vs Epoch Graph
Dice-Coefficient Vs Epoch Graph
Accuracy Vs Epoch Graph
As the data was available on kaggle we chose to use kaggle notebook as the data file was too large. Kaggle provides CPU and GPU as per requirement to test and train you model. Similar to other online compiler kaggle notebook is efficient and data is saved automatically.
[1] U.Baid, et al., "The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification", arXiv:2107.02314, 2021.
[2] B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, et al. "The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)", IEEE Transactions on Medical Imaging 34(10), 1993-2024 (2015) DOI: 10.1109/TMI.2014.2377694
[3] S. Bakas, H. Akbari, A. Sotiras, M. Bilello, M. Rozycki, J.S. Kirby, et al., "Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features", Nature Scientific Data, 4:170117 (2017) DOI: 10.1038/sdata.2017.117