Deep multi-scale feature fusion for pancreas segmentation from CT images

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Pancreas segmentation from computed tomography (CT) images is crucial for cancer detection and radiation therapy.
• Manual segmentation is labor-intensive and prone to variability.
• Existing automated methods struggle with diverse pancreas shapes and sizes.

Purpose of the Study:

• To develop an accurate and robust computer-assisted pancreas segmentation model.
• To address the limitations of current segmentation techniques in handling anatomical variations.
• To present the multi-scale feature fusion (MsFF) model for improved pancreas segmentation.

Main Methods:

• The proposed MsFF model utilizes an encoder-decoder framework.
• Incorporates a squeeze-and-excitation module in the encoder to enhance feature learning.
• Employs a hierarchical fusion module to integrate low-level and high-level features for boundary preservation.

Main Results:

• The MsFF model was evaluated on the NIH pancreas dataset.
• Achieved a Dice Sorensen Coefficient of 87.26%.
• Obtained a Volumetric Overlap Error of 22.67%, outperforming state-of-the-art methods.

Conclusions:

• The integration of squeeze-and-excitation and hierarchical fusion modules significantly enhances segmentation performance.
• The MsFF model demonstrates superior accuracy and robustness in pancreas segmentation.
• The findings support the efficacy of the proposed MsFF model for clinical applications.