Nbi-Infframes

The <strong>NBI-InfFrames </strong>dataset<strong> </strong>aims to provide the surgical data science community with a labeled dataset for the identification of informative endoscopic video frames. It is composed of 720 video frames. The frames are manually extracted and labeled from 18 narrow-band laryngoscopic videos of 18 different patients affected by laryngeal spinocellular carcinoma (diagnosed after histopathological examination). The frames include 180 informative (<strong>I</strong>) video frames, 180 blurred (<strong>B</strong>) frames, 180 frames with saliva or specular reflections (<strong>S</strong>) and 180 underexposed (<strong>U</strong>) frames. The dataset was created for testing the method proposed in S. Moccia, et al. "<em>Learning-based classification of informative laryngoscopic frames.</em>" COMPUTER METHODS AND PROGRAM IN BIOMEDICINE, (accepted for publication). The folder<em> <strong>FRAMES.zip</strong> </em>contains 3 subfolders (FOLD1, FOLD2, FOLD3), which are the 3 folds used for cross-validation purpose in the frame classification performance assessment. Data separation in the folds is performed both at image- and patient-level. Each subfolder contains 4 folders relative to the four frame classes, i.e., <strong>I</strong>, <strong>B</strong>, <strong>S</strong> and <strong>U</strong>.