Learning to recover sharp detail from simulated low-dose ct studies

Radiology exams require exposing a patient to a variable dosage of radiation. Importantly, the amount of radiation used during the exam directly corresponds to the level of noise in the resulting image, and increased amounts of radiation can pose health risks to patients. This results in a tradeoff, as radiologists need a high-quality image to make a diagnosis. In this work, we propose a method to recover image fidelity given a noisy, or low-dose, sample. Using a two-part criterion that consists of a pixel-wise loss and an adversarial loss, we are able to recover the structure and fine detail of the normal-dose sample. To evaluate the denoising method, we implement simulations of realistic low-dose noise for a computed tomography exam, which may be of independent interest. Quantitative and qualitative results highlight the performance of our approach as compared to existing baselines.