DyABD: the abdominal muscle segmentation in dynamic MRI benchmark

Area of Science:

• Medical Imaging
• Computer Vision
• Anatomy

Background:

• Abdominal hernias are common, with high recurrence rates impacting patient quality of life.
• Accurate abdominal muscle segmentation is crucial for understanding hernia pathophysiology and treatment effectiveness.
• Existing medical image datasets lack the dynamic variability and comprehensive annotations needed for robust model training.

Purpose of the Study:

• Introduce DyABD, a novel dynamic abdominal MRI dataset for abdominal muscle segmentation.
• Evaluate the generalization capabilities of current segmentation models on this challenging dataset.
• Establish a new benchmark for progress in medical image segmentation research.

Main Methods:

• Collected dynamic abdominal MRIs of patients performing exercises, capturing anatomical variability.
• Developed high-quality abdominal muscle annotations for the DyABD dataset.
• Assessed segmentation model performance across Supervised, Few Shot, and Zero Shot learning paradigms.

Main Results:

• DyABD presents a unique challenge due to extreme anatomical variability during dynamic exercises.
• Most existing segmentation models achieved a Dice Coefficient of approximately 0.82 on the DyABD dataset.
• Performance varied significantly across different learning paradigms, indicating generalization limitations.

Conclusions:

• The DyABD dataset provides a critical resource for advancing abdominal muscle segmentation research.
• Current medical image segmentation models require substantial improvement for reliable clinical application.
• DyABD redefines the benchmark for evaluating segmentation model progress in dynamic abdominal imaging.