Miguel Lopes Martins

Abstract
This paper is concerned with the semantic segmentation within domain-specific contexts, such as those pertaining to biology, physics, or material science. Under these circumstances, the objects of interest are often irregular and have fine structure, i.e., detail at arbitrarily small scales. Empirically, they are often understood as self-similar processes, a concept grounded in Multifractal Analysis. We find that this multifractal behaviour is carried out through a convolutional neural network (CNN), if we view its channel-wise responses as self-similar measures. A function of the local singularities of each measure we call Singularity Stregth Recalibration (SSR) is set forth to modulate the response at each layer of the CNN. SSR is a lightweight, plug-in module for CNNs. We observe that it improves a baseline U-Net in two biomedical tasks: skin lesion and colonic polyp segmentation, by an average of 1.36% and 1.12% Dice score, respectively. To the best of our knowledge, this is the first time multifractal-analysis is conducted end-to-end for semantic segmentation.

Abstract

Abstract
This work considers the problem of segmenting heart sounds into their fundamental components. We unify statistical and data-driven solutions by introducing Markov-based Neural Networks (MNNs), a hybrid end-toend framework that exploits Markov models as statistical inductive biases for an Artificial Neural Network (ANN) discriminator. We show that an MNN leveraging a simple onedimensional Convolutional ANN significantly outperforms two recent purely data-driven solutions for this task in two publicly available datasets: PhysioNet 2016 (Sensitivity: 0.947 +/- 0.02; Positive Predictive Value : 0.937 +/- 0.025) and the CirCor DigiScope 2022 (Sensitivity: 0.950 +/- 0.008; Positive Predictive Value: 0.943 +/- 0.012). We also propose a novel gradient-based unsupervised learning algorithm that effectively makes the MNN adaptive to unseen datum sampled from unknown distributions. We perform a cross dataset analysis and show that an MNN pre-trained in the CirCor DigiScope 2022 can benefit from an average improvement of 3.90% Positive Predictive Value on unseen observations from the PhysioNet 2016 dataset using this method.

Abstract
Gastric Intestinal Metaplasia (GIM) is a precancerous gastric lesion and its early detection facilitates patient followup, thus lowering significantly the risk of death by gastric cancer. However, effective screening of this condition is a very challenging task, resulting low intra and inter-observer concordance. Computer assisted diagnosis systems leveraging deep neural networks (DNNs) have emerged as a way to mitigate these ailments. Notwithstanding, these approaches typically require large datasets in order to learn invariance to the extreme variations typically present in Esophagogastroduodenoscopy (EGD) still frames, such as perspective, illumination, and scale. Hence, we propose to combine a priori information regarding texture characteristics of GIM with data-driven DNN solutions. In particular, we define two different models that treat pre-trained DNNs as general features extractors, whose pairwise interactions with a collection of highly invariant local texture descriptors grounded on fractal geometry are computed by means of an outer product in the embedding space. Our experiments show that these models outperform a baseline DNN by a significant margin over several metrics (e.g., area under the curve (AUC) 0.792 vs. 0.705) in a dataset comprised of EGD narrow-band images. Our best model measures double the positive likelihood ratio when compared to a baseline GIM detector.

Abstract
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 interfold 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.