Gliomas Autonomous Segmentation Using Advanced Deep Learning Techniques

Document Type : Primary Research paper

Authors

1 Bannari Amman Institute of Technology, Erode, Tamilnadu, India

2 VIT Bhopal University, Madhya Pradesh, India

3 PSG College of Technology, Coimbatore, Tamilnadu, India

4 Excel College of Engineering and Technology, Namakkal, Tamilnadu, India

Abstract

Segmentation of brain Gliomas from MRI autonomously is one of the important
tasks for accurate diagnosis and efficient treatment procedures. Lot of recent researches
involves many deep learning models for predicting the results in a proficient manner. Some
researches include Convolutional Neural Networks, both 2D and 3D approach. But results
gained through CNN are not promising and it is time and memory consuming. In our
proposed work, a com- puter vision based package FastAI based Dynamic_UNet model is
utilized. The model is fine_tuned by comparing the results with Classical_UNet. The model
results are visualized and the optimized parameters are chosen for segmentation process
using Neptune AI logger tool. The accuracy obtained is nearly close to the ground truth
results. The loss obtained is less than 0.005% with an accuracy of more than 87%. The
model results helps to overcome the uncertainty infor- mation obtained due to falsesegmentations
and helps to perk up the prediction accuracy.

Keywords