We propose two attention functions that capture cross-channel statistical scaling regularities: Monofractal and Multifractal recalibration. We build an experimental framework centered around the U-Net [1] and show that these approaches, especially Multifractal recalibration, lead to substantial improvements over a baseline augmented with other attention functions that may also describe each channel in terms of higher-order statistics [2]-[4]. Our experiments cover three public datasets from diverse modalities: ISIC18 (dermoscopy) [5], Kvasir-SEG (endoscopy) [6], and BUSI (ultrasound) [7]. Additionally, we also study the dynamics of the Squeezeand-Excite attention layer [8] and our findings suggest that (a) excitation response does not get increasingly specialized with encoder depth in the U-Net due to its skip connections, and that its effectiveness may be linked to global statistics of their instance-variability. To the best of our knowledge, we present the first instance of end-to-end multifractal analysis for semantic segmentation.

Gastric Intestinal Metaplasia (GIM) is one of the precancerous conditions in the gastric carcinogenesis cascade and its optical diagnosis during endoscopic screening is challenging even for seasoned endoscopists. Several solutions leveraging pre-trained deep neural networks (DNNs) have been recently proposed in order to assist human diagnosis. In this paper, we present a comparative study of these architectures in a new dataset containing GIM and non-GIM Narrow-band imaging still frames. We find that the surveyed DNNs perform remarkably well on average, but still measure sizeable inter- fold variability during cross-validation. An additional ad-hoc analysis suggests that these baseline architectures may not perform equally well at all scales when diagnosing GIM.