Power-law spectrum-based objective function to train a generative adversarial network with transfer learning for the...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Generative Adversarial Networks (GANs) show promise for synthesizing medical images.
• Existing GANs require optimization for specific medical imaging tasks like breast CT synthesis.
• X-ray-based breast images exhibit power-law spectral characteristics, a property not always leveraged in GAN training.

Purpose of the Study:

• To introduce a novel objective function, the beta loss function, to enhance breast CT image synthesis quality using GANs.
• To compare the efficacy of different transfer learning datasets (ImageNet vs. anatomical noise) for GAN-based breast CT image generation.
• To evaluate the performance of GANs incorporating the beta loss function on a patient-derived breast CT dataset.

Main Methods:

• Developed the beta loss function based on the power-law spectrum (beta value) of breast CT images, aiming for a beta value close to two.
• Integrated the beta loss function into the StyleGAN2 architecture.
• Utilized a patient-derived breast CT dataset (7355 training, 212 validation images) and compared transfer learning with ImageNet and anatomical noise datasets.

Main Results:

• The beta loss function resulted in synthetic images with beta values closer to real images and lower Fréchet Inception Distance (FID) scores.
• GANs pre-trained with anatomical noise images outperformed those pre-trained with ImageNet in both beta value evaluation and FID score.
• The combination of the beta loss function and anatomical noise transfer learning dataset achieved the lowest FID score.

Conclusions:

• The proposed beta loss function significantly improves GAN-based breast CT image synthesis.
• Anatomical noise images serve as a more effective transfer learning dataset than ImageNet for this application.
• The optimized GAN approach holds potential for advancing medical imaging applications through improved synthetic data generation.