Generation of multimodal realistic computational phantoms as a test-bed for validating deep learning-based...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Validating deep learning models for medical image translation requires high-quality, paired datasets, which are currently scarce.
• Existing methods for generating synthetic CT (sCT) from MRI face challenges in robust validation due to data limitations.

Purpose of the Study:

• To develop a novel framework using computational phantoms to generate realistic, paired CT and MRI images for AI model validation.
• To enable reliable ground-truth datasets for artificial intelligence (AI) methods generating synthetic CT (sCT) from MRI, specifically for radiotherapy.

Main Methods:

• Leveraged computational phantoms and CycleGANs (cycle-consistent generative adversarial networks) to create synthetic CT and MRI images with realistic textures.
• Trained generative adversarial networks (GANs) to transfer patient imaging styles onto phantoms.
• Evaluated synthetic data through paired and unpaired comparisons, dosimetric analysis, and external validation on public CT datasets.

Main Results:

• Generated synthetic CT/MRI phantoms demonstrated strong anatomical consistency with original phantoms.
• Achieved high histogram correlation with patient images (HistCC = 0.998 ± 0.001 for MRI, 0.97 ± 0.04 for CT).
• Validated a GAN-based model for sCT generation, showing dosimetric accuracy comparable to real data.

Conclusions:

• The proposed framework using generated phantoms provides a reliable method for validating deep learning-based cross-modality synthesis techniques.
• This novel approach addresses the critical need for robust validation datasets in AI-driven medical image translation for radiotherapy.