Improving the slice interaction of 2.5D CNN for automatic pancreas segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Manual pancreas segmentation is time-consuming and labor-intensive.
• Accurate pancreas segmentation is crucial for diagnosing pancreatic diseases, researching diabetes, and surgical planning.
• Deep learning offers a potential solution for automating this process.

Purpose of the Study:

• To develop a deep learning-based framework for automatic pancreas segmentation in 3D medical images.
• To improve the efficiency and accuracy of pancreas delineation compared to manual methods.
• To provide a tool for enhanced diagnosis and research related to pancreatic conditions.

Main Methods:

• A two-stage deep learning framework was designed, incorporating a Square Root Dice loss for improved localization.
• A novel 2.5D slice interaction network with a slice correlation module was developed for capturing cross-slice information.
• Self-supervised learning (slice shuffle), ensemble learning, and recurrent refinement were employed to enhance segmentation accuracy and robustness.

Main Results:

• The framework achieved state-of-the-art performance on the public NIH Pancreas-CT dataset.
• Dice similarity coefficient, sensitivity, and specificity were reported as 86.21±4.37%, 87.49±6.38%, and 85.11±6.49%, respectively.
• Fourfold cross-validation demonstrated the method's capability and robustness on 82 contrast-enhanced 3D CT scans.

Conclusions:

• The proposed automatic pancreas segmentation framework is effective and validated on an open dataset.
• The 2.5D network architecture benefits from multi-level slice interaction.
• Self-supervised pre-training significantly boosts neural network performance, offering potential for routine pancreatic disease diagnosis.