Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Graphical Approach For Brain Haemorrhage Segmentation

Ninad Mehendale, Pragya Gupta, Nishant Rajadhyaksha, Ansh Dagha, Mihir Hundiwala, Aditi Paretkar, Sakshi Chavan, Tanmay Mishra

TL;DR

This work tackles automated brain haemorrhage segmentation in CT scans under limited data. It introduces a novel three-module architecture combining a CNN-based encoder/decoder with a Self Constructing Graph (SCG) and Graph Convolutional Network (GCN) to capture both local features and global image structure, integrated via an inference module. The study demonstrates that a Dice loss combined with BCE loss yields the best Dice score of $0.81$, with the model achieving competitive performance while using fewer parameters than a baseline U-Net. The approach offers a data-efficient pathway for accurate haemorrhage delineation, with potential clinical impact in speeding up and improving treatment decisions in neurological emergencies.

Abstract

Haemorrhaging of the brain is the leading cause of death in people between the ages of 15 and 24 and the third leading cause of death in people older than that. Computed tomography (CT) is an imaging modality used to diagnose neurological emergencies, including stroke and traumatic brain injury. Recent advances in Deep Learning and Image Processing have utilised different modalities like CT scans to help automate the detection and segmentation of brain haemorrhage occurrences. In this paper, we propose a novel implementation of an architecture consisting of traditional Convolutional Neural Networks(CNN) along with Graph Neural Networks(GNN) to produce a holistic model for the task of brain haemorrhage segmentation.GNNs work on the principle of neighbourhood aggregation thus providing a reliable estimate of global structures present in images. GNNs work with few layers thus in turn requiring fewer parameters to work with. We were able to achieve a dice coefficient score of around 0.81 with limited data with our implementation.

A Graphical Approach For Brain Haemorrhage Segmentation

TL;DR

This work tackles automated brain haemorrhage segmentation in CT scans under limited data. It introduces a novel three-module architecture combining a CNN-based encoder/decoder with a Self Constructing Graph (SCG) and Graph Convolutional Network (GCN) to capture both local features and global image structure, integrated via an inference module. The study demonstrates that a Dice loss combined with BCE loss yields the best Dice score of , with the model achieving competitive performance while using fewer parameters than a baseline U-Net. The approach offers a data-efficient pathway for accurate haemorrhage delineation, with potential clinical impact in speeding up and improving treatment decisions in neurological emergencies.

Abstract

Haemorrhaging of the brain is the leading cause of death in people between the ages of 15 and 24 and the third leading cause of death in people older than that. Computed tomography (CT) is an imaging modality used to diagnose neurological emergencies, including stroke and traumatic brain injury. Recent advances in Deep Learning and Image Processing have utilised different modalities like CT scans to help automate the detection and segmentation of brain haemorrhage occurrences. In this paper, we propose a novel implementation of an architecture consisting of traditional Convolutional Neural Networks(CNN) along with Graph Neural Networks(GNN) to produce a holistic model for the task of brain haemorrhage segmentation.GNNs work on the principle of neighbourhood aggregation thus providing a reliable estimate of global structures present in images. GNNs work with few layers thus in turn requiring fewer parameters to work with. We were able to achieve a dice coefficient score of around 0.81 with limited data with our implementation.
Paper Structure (18 sections, 15 equations, 6 figures, 3 tables)

This paper contains 18 sections, 15 equations, 6 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Methodology
  4. Dataset
  5. Model Architecture
  6. Convolution Neural Network Module
  7. Graph Neural Network Module
  8. SCG Network
  9. GCN Layer
  10. Inference Module
  11. Results and Discussions
  12. Objective Functions
  13. Dice Loss
  14. Binary Cross Entropy Loss
  15. Focal Tversky Loss
  16. ...and 3 more sections

Figures (6)

  • Figure 1: Images of Brain haemorrhage CT-Scan Slices
  • Figure 2: Images of Brain haemorrhage CT-Scan Segmentation Masks
  • Figure 3: Model Architecture
  • Figure 4: Self Constructing Graph Network
  • Figure 5: Validation Dice Score V/s Training Steps
  • ...and 1 more figures