Enhance Precision in Glioblastoma Multiforme Resection: Tumor Localization Technique

Overview

Background Glioblastoma Multiforme (GBM) is a highly aggressive and deadly form of brain cancer that not only currently has a poor prognosis but an extremely low survival duration after surgery is conducted. The complexity of surgical procedures and resection related to GBM is heightened by the difficulty in distinguishing tumor margins. Every year over 12,000 cases emerge and very often even though surgeries are successfully completed, the median survival after-the-fact is only a mere 15 months and with only a mere 5% chance for a five-year survival rate. Methods and Findings In our research, to train an initial model, one that is focused on identifying areas of tumor presence and areas of non-tumor presence by incorporating annotated and segmented MRI data for GBM such as the Upenn BRaTS or LUMIERE datasets. The U-net architecture is first constructed by preprocessing all types of MRIs, CT1, FLAIR, T2, and T1 into 256x256 images with 32 slices in 3x3 kernels, then these standardized images are put through a rigorous downsampling phase to a image resolution of 16x16 with 32 slices. Then upsampled from 16x16 to 256x256 while still withholding the 32 slices. More specifically the U-net architecture is built through a series of Convolutional Neural Network systems specifically using techniques such as batch Normalization, Pooling and Filtering in order to aid the model in analyzing MRI images. To seek a trained initial model for basic sigmoid tumor identification, we first analyzed a multitude of modalities from a dataset dedicated to GBM MRIs. Additionally, we also used the aid of masks for each individual MRI image in order to create a more accurate U-Net model. In the end we created a model with a training accuracy of 98.131% with a loss of only 7.119%. Additionally we validated that overfitting and false positives were not appearing as an evaluation accuracy was 98.206% and 7.769% for evaluation loss meaning. Additionally utilizing this model in a HDF5, hierarchical data format file, we built a second mask generation model that specifically aimed to highlight non-tumor areas in order to single out the location of tumorous areas. Then we attempted to rerun the experiment using two GBM MRI datasets, specifically now adding the BRaTS UPenn Dataset and we achieved extremely similar results showing us that our U-net model is resilient and more importantly extremely cross compatible. Conclusion Utilizing two different datasets for GBM MRIs and masks specifically the LUMIERE Dataset and the UPenn BRaTS dataset in different modalities Ct1, T1, T2, and FLAIR, we trained a initial model to identify Gliomas in the brain . Then we used the initial modelâ€™s HDF5 files in order to create a masking technique to identify and highlight non-tumor areas to improve safety and efficacy of GBM resection.