Slice imputation: Multiple intermediate slices interpolation for anisotropic 3D medical image segmentation.

Area of Science:

• Medical Imaging
• Image Processing
• Computer Vision

Background:

• Anisotropic 3D medical images present challenges for accurate segmentation due to uneven slice resolution.
• Existing inter-slice imputation methods often lack comprehensive smoothness considerations across all spatial dimensions.

Purpose of the Study:

• To develop a novel frame-interpolation method for slice imputation to enhance segmentation accuracy in anisotropic 3D medical images.
• To improve the smoothness of interpolated 3D medical volumes in axial, sagittal, and coronal directions.

Main Methods:

• Introduced a frame-interpolation-based method for slice imputation in anisotropic 3D medical volumes.
• Proposed a multitask inter-slice imputation approach incorporating a smoothness loss function for through-plane direction evaluation (sagittal and coronal).
• Transformed interpolated volumes into isotropic representations to improve segmentation performance.

Main Results:

• The method improved resolution in the through-plane direction and achieved isotropic representations.
• Demonstrated superior performance over competing slice imputation methods in whole brain tumor, liver tumor, and prostate segmentation.
• Achieved up to 1% Dice improvement for CT liver tumor segmentation and over 2% Dice improvement for MRI prostate segmentation.

Conclusions:

• The proposed method effectively enhances segmentation accuracy for anisotropic 3D medical images.
• The frame-interpolation technique with multi-directional smoothness optimization offers significant improvements in medical image analysis.
• This approach holds promise for improving diagnostic accuracy in various clinical applications using CT and MRI data.