Medical Object Detection

Object localization and categorization in medical images is typically formulated as a problem of semantic segmentation. This fails to directly tackle the coarse localization task by learning pixel-level scores, but requires ad-hoc heuristics during back-transformation to object-level scores. Current state-of-the-art object detectors, in contrast, allow individual object scoring end-to-end, ironically giving up the ability to tap the full pixel-wise supervision signal. we present Retina U-Net, a simple architecture, which naturally merges the Retina Net one-stage detector with the U-Net architecture widely used for semantic segmentation in medical images. We look at the importance of full segmentation monitoring on two health data sets, provide a detailed study on a series of toy experiments and show how the equivalent performance gain grows in the limit of very small data sets.