Robust Blood Cell Image Segmentation Method Based on Neural Ordinary Differential Equations.

Area of Science:

• Medical image analysis
• Computational biology
• Artificial intelligence in healthcare

Background:

• Accurate segmentation of blood cell images is crucial for disease diagnosis via microscopy.
• Current methods require precise identification of cell morphology, number, and ratios.
• Existing segmentation techniques may face limitations in accuracy and computational efficiency.

Purpose of the Study:

• To enhance blood smear image segmentation accuracy using a combination of neural ordinary differential equations (NODEs) and U-Net networks.
• To investigate the impact of integrating an ODE-block module on network speed and accuracy.
• To optimize the NODEs-UNet model for improved cell counting and identification in medical imaging.

Main Methods:

• Blood cell images were preprocessed to improve contrast for segmentation.
• An ODE-block module was integrated into nine convolutional layers of the U-Net architecture.
• The NODEs-UNet model was trained and tested on a unified dataset, with error tolerance adjusted for optimal performance.
• Experimental results guided the selection of ODE-block placement and error tolerance to balance speed and accuracy.

Main Results:

• The proposed NODEs-UNet model achieved 95.3% pixel accuracy and 90.61% mean intersection over union on the testing set.
• Compared to standard U-Net and ResNet, the NODEs-UNet model showed improvements in pixel accuracy (0.88% and 0.46%) and mean intersection over union (2.18% and 1.13%).
• The integration of ODE-blocks optimized computational cost without compromising segmentation accuracy.

Conclusions:

• The NODEs-UNet model significantly improves the accuracy of blood cell image segmentation.
• This approach offers a more efficient computational method for medical image analysis.
• The findings suggest a promising direction for advancing automated disease diagnosis through enhanced cell image segmentation.