Related Experiment Video
Updated: Jul 18, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Bridged-U-Net-ASPP-EVO and Deep Learning Optimization for Brain Tumor Segmentation
Rammah Yousef1, Shakir Khan2,3, Gaurav Gupta1
1Yogananda School of AI, Computers and Data Sciences, Shoolini University, Solan 173229, India.
Diagnostics (Basel, Switzerland)
|August 26, 2023
Summary
This study enhances brain tumor segmentation in MRI using deep learning, introducing a novel Bridged U-Net-ASPP-EVO model. The new architecture significantly improves segmentation accuracy for various tumor sub-regions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuro-oncology
Background:
- Manual segmentation of brain tumors in MRI is complex and time-consuming for radiologists.
- Accurate segmentation is crucial for diagnosis, treatment planning, and monitoring of brain tumors.
- Deep learning offers potential for automating and improving the accuracy of brain tumor segmentation.
Purpose of the Study:
- To investigate the impact of deep learning optimizers and loss functions on brain tumor segmentation.
- To introduce and evaluate a novel deep learning architecture, Bridged U-Net-ASPP-EVO, for enhanced brain tumor segmentation.
- To compare the performance of the proposed model against state-of-the-art methods on benchmark datasets.
Main Methods:
- Experimental evaluation of deep learning optimizers and loss functions for brain tumor segmentation.
- Development of the Bridged U-Net-ASPP-EVO architecture incorporating Atrous Spatial Pyramid Pooling, Evolving Normalization, squeeze and excitation blocks, and max-average pooling.
- Validation of two variants (v1 and v2) of the proposed architecture on the MICCAI BraTS 2020 and RSNA-ASNR-MICCAI BraTS 2021 datasets.
Main Results:
- The Bridged U-Net-ASPP-EVO models achieved competitive results compared to existing state-of-the-art models.
- Achieved average segmentation Dice scores of 0.84, 0.85, 0.91 for variant 1 and 0.83, 0.86, 0.92 for variant 2 on the BraTS 2021 validation dataset for ET, TC, and WT sub-regions, respectively.
- Demonstrated the effectiveness of incorporating multi-scale information processing and advanced normalization techniques for improved segmentation.
Conclusions:
- The proposed Bridged U-Net-ASPP-EVO architecture effectively segments brain tumors from MRI scans.
- The study highlights the importance of architectural components like Atrous Spatial Pyramid Pooling for handling diverse tumor sizes.
- The developed models show significant promise for clinical application in automated brain tumor segmentation.
