Super-resolution reconstruction improves multishell diffusion: using radiomics to predict adult-type diffuse glioma...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Multishell diffusion scanning offers rich information but is limited by low spatial resolution.
• Deep learning-based super-resolution reconstruction (SR) is a promising technique to overcome these limitations.

Purpose of the Study:

• To enhance the spatial resolution of multishell diffusion images using SR.
• To develop and validate prediction models for adult-type diffuse glioma, including isocitrate dehydrogenase (IDH) status and tumor grade (2/3).

Main Methods:

• Diffusion tensor imaging (DTI), DKI, MAP, and NODDI models were constructed from multishell diffusion data.
• A generative adversarial network (GAN) based on deep residual channel attention networks was used for SR, generating 2x and 4x resolution-improved images.
• Radiomic features were extracted and used to build diagnostic models via multiple pipelines.

Main Results:

• Visually, both 2x and 4x SR-improved images were superior to original images; 2x-improved images yielded better predictions.
• Advanced diffusion models (DKI, MAP, NODDI) did not outperform the simple DTI model in diagnostic performance.
• The NODDI model with 2x SR achieved the highest performance for IDH status prediction (AUC=0.877).
• The MAP model using original images was best for grading 2/3 tumors (AUC=0.806).

Conclusions:

• SR significantly improves multishell diffusion image resolution.
• The optimal SR resolution and diffusion model depend on the specific clinical prediction task.
• SR offers distinct advantages for different diagnostic goals in glioma analysis.