A preprocessing method based on 3D U-Net for abdomen segmentation.

Area of Science:

• Biomedical imaging
• Medical image analysis
• Deep learning

Background:

• Deep learning excels in biomedical automatic segmentation but requires optimization.
• Preprocessing methods are crucial for improving segmentation performance and speed.
• 3D U-Net is a proven architecture for segmentation tasks.

Purpose of the Study:

• To propose and evaluate a preprocessing step for enhanced biomedical image segmentation.
• To improve the accuracy and speed of abdomen region of interest (ROI) segmentation using 3D U-Net.
• To assess the impact of various training parameters and loss functions on segmentation outcomes.

Main Methods:

• Training a 3D U-Net model on combined CHAOS and AbdomenCT-1K datasets (6998 slices).
• Testing the model's generalizability on the AbdomenCT-1K dataset (1311 slices).
• Systematic examination of k-fold cross-validation (CV), batch sizes (bs), learning rates (lr), and loss functions (Dice, Focal Dice, Focal Twersky).
• Evaluation of segmentation performance using Dice score, Hausdorff Distance (HD), HD95, and Average Symmetric Surface Distance (ASSD).
• Utilizing Connected Components Analysis (CCA) for abdomen ROI identification and dimensionality reduction.

Main Results:

• The best recorded Dice score reached 99.71%.
• Fine-tuning parameters and loss functions significantly impacted segmentation performance.
• Connected Components Analysis (CCA) achieved an average dimensionality reduction of 33.34% for the test dataset.
• The proposed preprocessing step demonstrated effectiveness in improving segmentation.

Conclusions:

• The integration of a preprocessing step significantly enhances 3D U-Net performance for biomedical image segmentation.
• The study highlights the importance of parameter tuning and loss function selection for optimal results.
• The method effectively identifies the abdomen ROI and reduces data dimensionality, paving the way for more efficient analysis.