Related Experiment Video
Updated: May 15, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
Benchmark of Deep Encoder-Decoder Architectures for Head and Neck Tumor Segmentation in Magnetic Resonance Images:
1Department of Measurement and Electronics, AGH University of Krakow, Krakow, Poland.
Head and Neck Tumor Segmentation for Mr-Guided Applications : First MICCAI Challenge, HNTS-MRG 2024, Held in Conjunction with MICCAI 2024, Marrakesh, Morocco, October 17, 2024, Proceedings
|April 9, 2025
Summary
A traditional residual UNet-based method outperformed newer deep learning models for segmenting head and neck cancer in MRI scans, highlighting the importance of data preparation and preprocessing in radiation therapy planning.
Area of Science:
- Medical imaging and radiation oncology
- Artificial intelligence in healthcare
- Deep learning for medical image segmentation
Background:
- MRI-guided radiation therapy is increasingly used for head and neck cancer due to its imaging advantages.
- Accurate segmentation of tumors in MRI is crucial for effective radiation therapy planning and delivery.
- Manual segmentation is time-consuming and challenging, motivating the need for automated methods.
Purpose of the Study:
- To benchmark state-of-the-art deep learning segmentation architectures for head and neck cancer in MRI.
- To evaluate the impact of advanced encoder-decoder architectures in low-data, low-contrast scenarios.
- To identify the most effective methods for pre- and mid-therapy tumor segmentation.
Main Methods:
- Benchmarking of several deep learning segmentation architectures, including UNETR, SwinUNETR, SegMamba, and a residual UNet-based method.
- Evaluation on the HNTSMRG challenge dataset for head and neck cancer MRI segmentation.
- Comparison of segmentation performance using Dice Score (DSC) metrics.
Main Results:
- The residual UNet-based method achieved superior performance (DSC = 0.775/0.701) compared to newer architectures like UNETR (DSC = 0.617/0.657), SwinUNETR (DSC = 0.757/0.700), and SegMamba (DSC = 0.708/0.683).
- The lWM team's method (residual UNet-based) achieved a mean aggregated Dice score of 0.771 (pre-therapy) and 0.707 (mid-therapy).
- Results indicate that data preparation, objective function, and preprocessing are more critical than network architecture for this specific task.
Conclusions:
- Traditional methods like residual UNet can be highly effective for challenging medical image segmentation tasks, even with limited data and low contrast.
- Data preparation and preprocessing strategies play a more significant role than advanced network architectures in achieving accurate head and neck cancer segmentation in MRI.
- Further research should focus on optimizing these foundational elements for improved automated segmentation in radiation oncology.
