I³Net: Inter-Intra-Slice Interpolation Network for Medical Slice Synthesis
IEEE Transactions on Medical Imaging
|April 26, 2024
Summary
Medical imaging reconstruction is improved by a novel network that enhances low-resolution through-plane data by interpolating axial slices. This method achieves superior results compared to existing super-resolution techniques.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Medical imaging techniques like CT and MRI often produce anisotropic data with high in-plane and low through-plane resolution due to thicker slice reconstruction.
- This anisotropy limits the quality and detail of 3D medical image volumes.
Purpose of the Study:
- To address the limitations of anisotropic medical imaging data.
- To propose a novel deep learning network for enhancing through-plane resolution in medical images.
Main Methods:
- Developed the Inter-Intra-slice Interpolation Network (ISINet) leveraging axial slice interpolation for improved through-plane resolution.
- Incorporated a through-plane branch for supplementing limited resolution data and an in-plane branch for global frequency domain feature learning.
- Introduced a cross-view block to integrate information from all three imaging views.
Main Results:
- ISINet significantly outperforms state-of-the-art methods in super-resolution, video frame interpolation, and slice interpolation.
- Achieved a peak signal-to-noise ratio (PSNR) of 43.90dB, with over 1.14dB improvement at a ×2 upscale factor on the MSD dataset.
- Demonstrated faster inference times compared to existing methods.
Conclusions:
- The proposed ISINet effectively compensates for low through-plane resolution in anisotropic medical imaging data.
- Slice-wise interpolation from the axial view offers greater benefits than traditional super-resolution approaches for this data type.
- ISINet represents a significant advancement in medical image reconstruction, offering improved quality and efficiency.
