Many applications in machine vision and medical imaging require the capture of images from a scene with very low radiance, which may result in very noisy images and videos. An important example of such an application is the imaging of fluorescently-labeled tissue in fluorescence-guided surgery. Medical imaging systems, especially when intended to be used in surgery, are designed to operate in well-lit environments and use optical filters, time division, or other strategies that allow the simultaneous capture of low radiance fluorescence video and a well-lit visible light video of the scene. This work demonstrates video denoising can be dramatically improved by utilizing deep learning together with motion and textural cues from the noise-free video.

The following passage from the main article explains the dataset capture: "We capture data with a transient lighting-enabled, clinical wide-field FGS imaging system (OnLume Surgical, Madison, WI). OnLume’s system uses two camera sensors for both the reference and fluorescent cameras. In a surgical training model, we inject the near-infrared fluorescent agent, ICG, via syringe into the femoral artery of four chicken thighs to simulate vascular surgery. We prepared varying doses of ICG up to the clinical guidelines of 2.5 mg/mL to generate fluorescent videos with visual contrast and low noise that can be treated as ground truth and simulate much lower fluorescence. In future work, we would like to detect markers that have much fewer photons than our captured fluorescent videos. We capture about 100 minutes of simulated surgical footage with a variety of motion such as cutting, pulling, squeezing, injecting, and working with surgical tools. The 100 minutes of footage is broken up into 590 100-frame long videos. The videos are captured at 15 frames per second at a resolution of 768 × 1024."