Accurate pancreas segmentation using multi-level pyramidal pooling residual U-Net with adversarial mechanism

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Accurate pancreas segmentation is crucial for diagnosing and treating pancreatic diseases.
• Existing segmentation methods struggle with the complexity of medical imaging datasets like the NIH Pancreas-CT.
• Developing robust and precise segmentation algorithms remains a significant challenge in medical image analysis.

Purpose of the Study:

• To introduce a novel multi-level pyramidal pooling residual U-Net with an adversarial mechanism for improved organ segmentation.
• To evaluate the proposed model's performance on the challenging NIH Pancreas-CT dataset for pancreas segmentation.
• To enhance gradient flow and contextual information gathering for more accurate medical image segmentation.

Main Methods:

• Implemented a residual learning approach within an adversarial U-Net architecture to optimize gradient flow.
• Introduced a multi-level pyramidal pooling module (MLPP) to effectively gather contextual information for segmentation.
• Utilized four-fold cross-validation on 82 pancreatic contrast-enhanced abdominal CT volumes to assess model performance using Dice Similarity Coefficient (DSC) and recall.

Main Results:

• The proposed method outperformed the baseline network by 5.30% in DSC and 6.16% in recall.
• Achieved competitive results compared to state-of-the-art segmentation methods.
• Demonstrated robust performance across different pooling block configurations within the MLPP.

Conclusions:

• The developed algorithm exhibits excellent pancreas segmentation performance on a challenging dataset.
• The novel U-Net architecture with adversarial and MLPP mechanisms proves to be a satisfactory and promising tool for medical image segmentation.
• The findings suggest potential for broader application in clinical settings requiring precise organ segmentation.