Untapped Brain

International Live Conference on

August 05, 2023
Live Conference

Sessions:
 

  • Untapped Brain Potential

  • Comparative Neuroanatomy

  • Brain Study Technology and Experiments

  • LinkedIn
  • Twitter
  • Instagram
  • Facebook
  • YouTube
Dr Tshetiz Dahal  Lugansk State Medical University Ukraine
Dr Tshetiz Dahal Lugansk State Medical University Ukraine

Deep Spatio-Spatial models for Classifying Brain Tumors in MR Images

press to zoom
Dr Tshetiz Dahal  Lugansk State Medical University Ukraine
Dr Tshetiz Dahal Lugansk State Medical University Ukraine

Deep Spatio-Spatial models for Classifying Brain Tumors in MR Images

press to zoom
1/1

Deep Spatio-Spatial models for Classifying Brain Tumors in MR Images

 

A brain tumor is a mass or cluster of abnormal cells in the brain that has the potential to spread to other tissues nearby and pose a serious threat to the patient's life. For effective treatment planning, a precise diagnosis is necessary, and the main imaging technique for determining the extent of brain tumours is magnetic resonance imaging. The majority of this increase in Deep Learning techniques for computer vision applications may be attributed to the availability of a sizable

amount of data for model training and the advancements in model designs that produce better approximations in a supervised environment. The availability of free datasets with trustworthy annotations has significantly improved the classification of cancers using such deep learning techniques. These techniques often use either 3D models that employ 3D volumetric MRIs or even 2D models that take each slice into account separately. However, spatiotemporal models can be used as "spatio-spatial"; models for this job by treating each spatial dimension individually or by seeing the slices as a succession of images through time. These models can learn certain spatial and temporal correlations while using less processing power.This study classifies several types of brain tumours using two spatiotemporal models, ResNet (2+1)D and ResNet Mixed Convolution. It was found that both of these models outperformed ResNet18, a model that only used 3D convolutions. It was also shown that pre-training the models on a distinct, even unrelated dataset before training them for the objective of cancer classification enhances performance. As a result of these studies, Pre-trained ResNet Mixed Convolution was shown to be the most effective model, achieving a macro F1-score of 0.9345 and a test accuracy of 96.98% while also being the model with the lowest computing cost.

Dr. Tshetiz Dahal

Lugansk State Medical University

Ukraine