Rare lesions – accounting for less than 1 per cent of images – almost always escape detection by the models. By taking the rest of the video into account, the HCE method increases their recall by a factor of 2.7 to 3.1.
The endoscopic videocapsule explores the small intestine without intervention, but produces tens of thousands of images, which AI helps sort. Problem: Some clinically significant abnormalities appear on less than 1% of annotated images. The models then show good overall scores while almost entirely overlooking these rare cases.
Amit Kumar Patel and Debesh Jha propose not to judge each image in isolation. Their method of decoding, called HCE, adjusts the decision made on an image based on the evidence present in the rest of the video of the same patient, information that a frozen image cannot contain. It extends an image-by-image correction already optimized for the average F1 score on all classes.
Results from two data sets
Tested on Kvasir-Capsule and GALAR, the method multiplies by 2.7 to 3.1 the recall of rare classes on GALAR relative to a simple temporal smoothing, with established statistical significance. On Kvasir-Capsule, the gain goes up to a factor 3.4, without being statistically confirmed. Faced with optimal transport correction, HCE remains better in the four configurations tested, with smaller gains.
Read the full article on the original source
Go to the source →