Final project submission for BU CAS CS 523
Brain tumor segmentation via u-net neural networks
Dataset can be downloaded here: https://www.kaggle.com/mateuszbuda/lgg-mri-segmentation
We want to not only indicate the tumors that exist clearly in brain MR images but also indicate those which cannot be seen from eyes to discove the tumors as sson as possible for treatment.
The dataset contains brain MR images together with manual FLAIR abnormality segmentation masks. The images were obtained from The Cancer Imaging Archive (TCIA). They correspond to 110 patients included in The Cancer Genome Atlas (TCGA) lower-grade glioma collection with at least fluid-attenuated inversion recovery (FLAIR) sequence and genomic cluster data available. These patients comes from 5 different hospitals. Tumor genomic clusters and patient data is provided in data.csv file.
All images are provided in .tif format with 3 channels per image. For 101 cases, 3 sequences are available, i.e. pre-contrast, FLAIR, post-contrast (in this order of channels). For 9 cases, post-contrast sequence is missing and for 6 cases, pre-contrast sequence is missing. Missing sequences are replaced with FLAIR sequence to make all images 3-channel. Masks are binary, 1-channel images. They segment FLAIR abnormality present in the FLAIR sequence (available for all cases).
The dataset is organized into 110 folders named after case ID that contains information about source institution.
Neural network architectures:
U-Net: Convolutional Networks for Biomedical Image Segmentation: https://arxiv.org/pdf/1505.04597.pdf
Papers that used this dataset:
Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm https://arxiv.org/pdf/1906.03720.pdf
We want to first filter MR images that reflect no tumor and the tumors that are easily detected by brain MR images. We implement simple network layers of color difference detection identify.
Then, for those tumors that cannot be detected by human eyes, we plan to implement a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The general architecture consists of a contracting path to capture the picture of tumors and a symmetric expanding path that enables the localization of where the tumors are.
In many visual tasks, especially in biomedical image processing, the desired output should include localization, i.e., a class label is supposed to be assigned to each pixel to indicate where exists the tumor. Therefore, we use a network in a sliding-window setup to predict the class label of each pixel by providing a local region (patch) around that pixel. In this U-net network, the main idea is to supplement a usual contracting network by successive layers and replace the pooling operators to upsampling operators to increase the resolution of the output. To guarantee the localization, high resolution features from the contracting path should be combined with the upsampled output. In this way, we reinforce the pixels where hidden tumors represent and demonstrate more directly for human eyes to observe.
This way, we can provide an efficient process of identifying and segmenting brain tumors from patient MRI scans, assisting doctors in their diagnosis and providing better care to a greater amount of people.