Related Experiment Video
Updated: Aug 8, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Benchmarking deep learning architectures for hyperspectral in-vivo brain tumor segmentation
Guillermo Vazquez1, Domenico Ragusa2, Emanuele Torti2
1Research Center for Industrial Electronics and Multimodal Systems (CEIMM), Universidad Politécnica de Madrid (UPM), Calle Ramiro de Maeztu 7, Madrid, 28031, Spain.
Computer Methods and Programs in Biomedicine
|August 6, 2026
Summary
Convolutional models excel in hyperspectral medical image segmentation, offering high accuracy with compact architectures. This study benchmarks deep learning models for brain tumor segmentation using hyperspectral imaging (HSI).
Area of Science:
- Medical Image Analysis
- Deep Learning (DL)
- Hyperspectral Imaging (HSI)
Background:
- Deep Learning (DL) is crucial for medical image analysis, particularly in surgical guidance.
- Hyperspectral Imaging (HSI) shows promise for in-vivo brain tumor segmentation during surgery.
- Existing DL architectures for HSI are often developed outside healthcare, with limited comparative studies on medical data.
Purpose of the Study:
- To systematically benchmark principal DL architectures for medical HSI segmentation.
- To establish a taxonomy of DL architectures based on their mechanisms (convolutions, attention, hybrid).
- To evaluate model performance on the SLIMBRAIN and HELICoiD HSI datasets.
Main Methods:
- Benchmarking of DL architectures developed in the last five years.
- Utilized 25-band SLIMBRAIN and 128-band HELICoiD datasets.
- Classified architectures based on convolutional, attention, and hybrid mechanisms.
Main Results:
- Convolutional models achieved the highest performance, with F1-scores up to 65.08% and 91.13%.
- Hybrid and attention-based models followed in performance.
- Compact architectures like DBDA and RSSAN demonstrated excellent accuracy-efficiency trade-offs.
Conclusions:
- Depth-Wise (DW) convolutions and spatial-spectral feature separation are key mechanisms for medical HSI segmentation.
- High performance is achievable with compact models across varying spectral dimensions.
- Findings offer practical design insights for future HSI applications in neurosurgery and beyond.