Related Experiment Video
Updated: Dec 21, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
Fully automatic brain tumor segmentation with deep learning-based selective attention using overlapping patches and
Mostefa Ben Naceur1, Mohamed Akil2, Rachida Saouli3
1Gaspard Monge Computer Science Laboratory, Univ Gustave Eiffel, CNRS, ESIEE Paris, F-77454 Marne-la-Vallée, France; Smart Computer Sciences Laboratory, Computer Sciences Department, Exact.Sc, and SNL, University of Biskra, Algeria.
Medical Image Analysis
|May 18, 2020
Summary
A new Deep Convolutional Neural Networks (CNNs) model accurately segments brain tumors using selective attention. This AI model achieves state-of-the-art performance on MRI scans, outperforming radiologists.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuro-oncology
Background:
- Glioblastoma multiforme (GBM) segmentation is crucial for treatment planning and outcome prediction.
- Accurate segmentation of high- and low-grade GBM tumors from MRI is challenging due to tumor heterogeneity and complex boundaries.
- Current segmentation methods often require manual input, are time-consuming, and may lack consistency.
Purpose of the Study:
- To develop a fully automatic Deep Convolutional Neural Networks (CNNs) model for accurate segmentation of high- and low-grade Glioblastoma brain tumors.
- To enhance feature extraction from MRI images using a selective attention mechanism inspired by the human visual system.
- To address challenges of class-imbalance and spatial relationships in image patch analysis for improved segmentation.
Main Methods:
- Proposed a novel CNNs model incorporating a selective attention mechanism to maximize relevant feature extraction from MRI scans.
- Addressed class-imbalance through equal sampling of image patches and weighted cross-entropy loss.
- Investigated the impact of Overlapping Patches versus Adjacent Patches, finding overlapping patches superior for incorporating global and local context.
Main Results:
- The proposed End-to-End Deep Learning model achieved state-of-the-art performance on the BRATS-2018 dataset.
- Achieved median Dice scores of 0.90 (whole tumor), 0.83 (tumor core), and 0.83 (enhancing tumor), surpassing radiologist performance (74%-85%).
- The model demonstrated computational efficiency, segmenting the whole brain in an average of 12 seconds.
Conclusions:
- The developed Deep Learning model provides accurate and reliable automatic segmentation of Glioblastoma brain tumors.
- The selective attention mechanism and optimized patch handling significantly improve segmentation performance.
- The model's efficiency and accuracy make it suitable for research and clinical applications in neuro-oncology.
