Deep Learning Subtraction Angiography: Improved Generalizability with Transfer Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Vascular Imaging

Background:

• Digital subtraction angiography (DSA) is crucial for vascular imaging but can suffer from misalignment artifacts.
• Generating high-quality DSA images, especially from data with motion artifacts, remains a challenge.

Purpose of the Study:

• To evaluate the effectiveness of deep learning subtraction angiography (DLSA) in producing synthetic DSA images free from misalignment artifacts.
• To assess the generalizability of DLSA across different vascular beds and its performance on images with motion artifacts.

Main Methods:

• A deep neural network (pix2pix) was trained on a motion-free dataset of cerebral, hepatic, and splenic vasculature.
• The trained DLSA model was tested on datasets containing motion artifacts.
• Radiologists assessed image quality using a 5-grade Likert scale, with subgroup analyses for transfer learning and novel vasculature.

Main Results:

• DLSA generated synthetic DSA images with significantly fewer background artifacts compared to traditional methods (1.9 vs 3.5, P = .01).
• No significant difference was observed in foreground vascular detail between DLSA and traditional DSA (3.1 vs 3.3, P = .19).
• Transfer learning significantly improved the quality of DLSA-generated images (P < .001).

Conclusions:

• DLSA effectively generates synthetic angiograms without misalignment artifacts.
• The DLSA method demonstrates improved performance with transfer learning.
• DLSA shows reliable generalizability to new vascular types not seen during initial training.