Autoencoding low-resolution MRI for semantically smooth interpolation of anisotropic MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• High-resolution medical images aid analysis but are not always feasible to acquire.
• Conventional upsampling methods for creating high-resolution images from low-resolution data lack contextual information.
• Existing deep learning super-resolution methods are supervised, requiring high-resolution training data.

Purpose of the Study:

• To develop an unsupervised deep learning method for synthesizing intermediate medical image slices from low-resolution data.
• To enable semantic interpolation in the through-plane direction by exploiting autoencoder latent spaces.
• To overcome the limitations of supervised deep learning methods by eliminating the need for high-resolution training examples.

Main Methods:

• An unsupervised deep learning semantic interpolation approach was proposed.
• Autoencoder latent spaces were utilized for semantically smooth interpolation.
• Intermediate slices were synthesized by decoding convex combinations of latent space encodings from adjacent slices.
• A novel loss function was introduced to enforce semantic similarity by exploiting spatial relationships within image volumes.

Main Results:

• The method was trained and evaluated on cardiac cine, neonatal brain, and adult brain MRI scans.
• The proposed method significantly outperformed cubic B-spline interpolation in Structural Similarity Index Measure and Peak Signal-to-Noise Ratio (p<0.001).
• The unsupervised nature obviates the need for high-resolution training data.

Conclusions:

• The unsupervised deep learning semantic interpolation approach effectively synthesizes intermediate medical image slices.
• The method offers a viable solution for enhancing medical image resolution in clinical settings where high-resolution data is scarce.
• This technique advances medical image analysis by leveraging deep learning without the constraints of supervised learning.