Related Experiment Video
Updated: Aug 22, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Deep Learning Hybrid Techniques for Brain Tumor Segmentation
Khushboo Munir1, Fabrizio Frezza1, Antonello Rizzi1
1Department of Information Engineering, Electronics and Telecommunications (DIET), Sapienza University of Rome, Via Eudossiana 18, 00184 Rome, Italy.
Sensors (Basel, Switzerland)
|November 11, 2022
Summary
This study introduces advanced deep learning models for brain tumor detection using MRI scans. Novel architectures significantly improve segmentation accuracy, aiding in earlier and more consistent diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Medical images are crucial for diagnosing and monitoring diseases like brain tumors.
- Accurate segmentation of brain tumors from Magnetic Resonance (MR) images is essential for effective treatment planning.
- Current deep learning methods show promise in extracting tumor features for clinical diagnosis.
Purpose of the Study:
- To propose novel deep neural network architectures for automated brain tumor detection and segmentation.
- To enhance the performance and consistency of brain tumor identification from MR images.
- To compare the efficacy of proposed convolutional neural network and inception module-based architectures against baseline models.
Main Methods:
- Development and testing of new deep neural network architectures, including MI-Unet and Hybrid Unet, incorporating inception modules.
- Utilizing convolutional neural networks for feature extraction and segmentation of brain tumors in MR images.
- Comparative analysis of proposed architectures against a baseline Unet model using metrics like dice score, sensitivity, and specificity.
Main Results:
- MI-Unet demonstrated a 7.5% increase in dice score, 23.91% in sensitivity, and 7.09% in specificity compared to the baseline.
- Depth-wise separable MI-Unet showed improvements of 10.83% in dice score, 2.97% in sensitivity, and 12.72% in specificity.
- Depth-wise separable hybrid Unet achieved the highest performance gains: 15.45% in dice score, 20.56% in sensitivity, and 12.22% in specificity over the baseline.
Conclusions:
- The proposed deep neural network architectures, particularly depth-wise separable hybrid Unet, significantly enhance brain tumor segmentation accuracy.
- Automated screening procedures using these advanced models offer more robust and consistent identification of brain tumors.
- These findings support the clinical utility of deep learning for improved brain tumor diagnosis and patient management.
