MedSegBench: A comprehensive benchmark for medical image segmentation in diverse data modalities.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Existing benchmarks lack comprehensiveness across diverse imaging modalities and segmentation tasks.
• Variability in image quality and dataset imbalance pose significant challenges.

Purpose of the Study:

• To introduce MedSegBench, a comprehensive benchmark for evaluating deep learning models in medical image segmentation.
• To provide standardized datasets and a framework for fair comparison of segmentation algorithms.
• To facilitate the development of robust and generalizable medical image segmentation models.

Main Methods:

• Curated 35 datasets covering ultrasound, MRI, and X-ray modalities, totaling over 60,000 images.
• Implemented standardized train/validation/test splits addressing data variability and imbalance.
• Evaluated U-Net architecture with various encoders (ResNets, EfficientNet, DenseNet) for binary and multi-class segmentation (up to 19 classes).

Main Results:

• MedSegBench encompasses a wide range of medical imaging modalities and segmentation complexities.
• The benchmark facilitates reproducible research and direct comparison of deep learning models.
• Demonstrated the utility of MedSegBench in assessing segmentation algorithm performance.

Conclusions:

• MedSegBench is the most comprehensive benchmark for medical image segmentation to date.
• It serves as a valuable resource for advancing the field of medical image analysis.
• Publicly available datasets and code encourage further research and development.