Phase identification for dynamic CT enhancements with generative adversarial network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Dynamic contrast-enhanced computed tomography (CT) is crucial for diagnosis and vascular identification.
• Manual contrast phase recording in CT scans is prone to errors, leading to missing or mislabeled data.
• Accurate contrast phase identification is challenging due to variations in protocols, vascular dynamics, and metabolism.

Purpose of the Study:

• To develop an imaging-based contrast phase identification method for dynamic abdominal CT.
• To utilize a proposed adversarial learning framework for phase identification across five contrast phases.
• To overcome limitations of manual phase recording in clinical CT scans.

Main Methods:

• A generative adversarial network (GAN) was employed as a disentangled representation learning model.
• A low-dimensional common representation and class-specific codes were fused in the hidden layer to model contrast phases.
• The method was trained on 36,350 CT slices from 400 subjects and validated on 2,216 slices from 20 independent subjects.

Main Results:

• The proposed adversarial learning network achieved a correspondence of 0.93.
• It significantly outperformed VGG (0.59), ResNet50 (0.62), StarGAN (0.72), and 3DSE (0.90) in accuracy.
• Statistical analysis confirmed the significant improvement using a Stuart-Maxwell test on normalized multiclass confusion matrices.

Conclusions:

• Adversarial learning, particularly for the discriminator, effectively captures contrast information across different phases.
• The proposed discriminator, integrated into a disentangled network, demonstrates promising performance for automated contrast phase identification.
• This approach offers a robust solution for accurate phase identification in dynamic abdominal CT imaging.