The high volume and quality of apparent diffusion coefficient (ADC) data containing clinically significant (CS) prostate cancer (PCa) are critical for automated PCa detection with a high accuracy. However, ADC data of CS PCa is scarce and costly to obtain in practice. This paper proposes a novel Generative Adversarial Network (GAN), named StitchAD-GAN, for synthesizing high-quality ADC images of CS PCa. Our StitchADGAN employs a StitchLayer in the generator to address the difficult-to-optimize problem in most GANs. Instead of directly optimizing a complex generation from a low dimensional noise to an ADC image, we optimize n easier generations of sub-images, which are then aggregated into a full size image in the StitchLayer. Our discriminative module approximates two distances: 1) the Wasserstein distance (W-distance) between the synthetic and real ADC data of CS PCa, and 2) an auxiliary distance (AD) Jensen-Shannon divergence (JSD) between the synthetic CS PCa and real nonCS PCa data. By minimizing the W-distance and maximizing the JSD simultaneously, our StitchAD-GAN can capture CS PCa features in addition to predominant prostate gland information, and in turn synthesize more clinically meaningful ADC data of CS PCa. Visual and quantitative results demonstrate greater quality of our synthetic CS PCa data than those of the state-of-the-art methods and even real data.

Publications