Brazil

According to the World Health Organization, 8 million deaths are counted due to gastrointestinal diseases annually. Automatic detection of gastrointestinal landmarks is a task that can further help medical professionals reducing cost and time in exploratory exams. Computer-aided detection and diagnosis systems have been widely explored in the scientific field. However, it takes a lot of processing power to achieve satisfactory results. In order to overcome this problem, this work uses a simple Convolutional Neural Network together with the \textit{Triplet Loss} cost function to extract image characteristics of 3 gastrointestinal anatomical landmarks (z-line, pylorus, and cecum) to classify those images.For the training it&#39;s used the dataset Kvasir-v2, obtaining 96,60% of Precision, 97,7\% of Accuracy, 96,91% of Specificity, 98,61% of Recall 97,59% of F1-score.

BRACIS 2021

Classification of anatomical landmarks in the gastrointestinal tract with endoscopy images utilizing convolutional neural networks and triplet loss

anatomical landmarks

medical image classification

convolutional neural network

**BRACIS** originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence – **SBIA** (22 editions), and the Brazilian Symposium on Neural Networks – **SBRN** (13 editions). **ENIAC** is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. **STIL** is the bi-annual Language Technology event.
All events are supported by the Brazilian Computer Society – Special Interest Groups: **CE-IA**, **CE-IC** and **CE-PLN**.

To access the 10th Brazilian Conference on Intelligent Systems (**BRACIS**), you need to register for the conference.  Please visit the Conference official site to register

[Register now](http://c4ai.inova.usp.br/bracis2021/)

Register now

BRACIS originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence – SBIA (22 editions), and the Brazilian Symposium on Neural Networks – SBRN (13 editions). 
ENIAC is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. 
STIL is the bi-annual Language Technology event.
All events are supported by the Brazilian Computer Society – Special Interest Groups: CE-IA, CE-IC and CE-PLN

According to the World Health Organization, 8 million deaths are counted due to gastrointestinal diseases annually. Automatic detection of gastrointestinal landmarks is a task that can further help medical professionals reducing cost and time in exploratory exams. Computer-aided detection and diagnosis systems have been widely explored in the scientific field. However, it takes a lot of processing power to achieve satisfactory results. In order to overcome this problem, this work uses a simple Convolutional Neural Network together with the \textit{Triplet Loss} cost function to extract image characteristics of 3 gastrointestinal anatomical landmarks (z-line, pylorus, and cecum) to classify those images.For the training it's used the dataset Kvasir-v2, obtaining 96,60% of Precision, 97,7\% of Accuracy, 96,91% of Specificity, 98,61% of Recall 97,59% of F1-score.

technical paper

The core point of the research process are data. They are records fromscientific investigation, which support the results published in journals and con-ferences. Making research data available in open access digital repositories hasmany advantages, such as increasing the visibility of associated publications,reproducing experiments, and validating results. In Brazil, full and unrestrictedsharing of them is not yet accepted by most researchers. This paper presents aninitial study to describe a model analyzing the attitude of Brazilian researchersconcerning open access research data. A clustering algorithm was used to iden-tify different research profiles. The achieved results indicate the main reasonswhy the researchers object to share their data

A clustering algorithm to evaluate the attitude of Brazilian researchers regarding open access research data

Este trabalho propõem uma abordagem para auxiliar pessoas com deficiência visual no reconhecimento de pessoas independente da idade. O objetivo é desenvolver um sistema que utilize uma abordagem de reconhecimento facial, com foco na invariância na idade, que retorne bons resultados comparados aos resultados obtidos na revisão da literatura. A abordagem estudada utiliza Redes Neurais Convolucionais profundas - CCNs, pré-treinadas pelo conjunto de dados VGGFace2, para extrair descritores de características de imagens de faces e classificar com o algoritmo de classificação Linear SVM. Como pode ser visto no decorrer do trabalho, a abordagem retornou 89,9% de acurácia, utilizando o conjunto de dados FG-NET, com 1002 imagens. E utilizando o conjunto de dados CACD, que contém 163.446 imagens divididas em quatro subconjuntos diferentes, três conjuntos para treino e um para teste, a abordagem retornou 85,2%, 82,4% e 88,2% de acurácia para cada modelo treinado com um conjunto de treinamento diferente.

Age-Invariant Face Recognition to Aid Visually Impaired Pleople

The employment of video surveillance cameras by public safety agencies enables incident detection in monitored cities by using object detection for scene description, enhancing the protection to the general public. Object detection has its drawbacks, such as false positives. Our work aims to enhance object detection and image classification by employing IoU (Intersection over Union) to minimize the false positives and identify weapon holders or fire in a frame, adding more information to the scene.

Detection of weapon possession and fire in Public Safety surveillance cameras

Handwritten signature authentication systems are important in many real world scenarios to avoid frauds. 
Thanks to Deep Learning, state-of-art solutions have been proposed to this problem by making use of Convolutional Neural Networks, but other models in this Machine Learning subarea are still to be furtherexplored. 
In this perspective, the present article introduces a Conditional Deep
Convolutional Generative Adversarial Networks (cDCGAN) approach whose experimental results in a realistic dataset with skilled forgeries have Equal Error Rate (EER) of 18.53 % and balanced accuracy of 87.91 %. 
These results validate a writer-dependent cDCGAN-based solution to the signature authentication problem in a real world scenario where no forgeries are available nor required in training time.

Offline Handwritten Signature Authentication with Conditional Deep Convolutional Generative Adversarial Networks

Summarized the main findings of the paper Face Reconstruction with Variational Autoencoder and Face Masks, answering key questions of how the set up of Face Masks and combined loss functions between L1, L2, and SSIM affects the image reconstruction quality.

Face Reconstruction with Variational Autoencoder and Face Masks

Our work analyses a set of semantic segmentation methods applied todetect people on beach images, as part of an automatic track system to monitorbathers to prevent trespassing the boundaries of the safe region for swimming.In our analysis, we compared the semantic segmentation networks U-net, X-net,LinkNet and Unet++ with the pretrained backbones VGG-16 and VGG-19. Webuilt our own dataset, composed of 300 images. The models were evaluatedbound box-wise using F-score metrics. Our findings show that X-Net achievedthe best value for F-score, with 90.89%, while Linknet was faster than the othernetworks, with no significant statistical difference in the F-score values.

Semantic Segmentation for People Detection on Beach Images

Image classification has been applied to several real problems.
However, getting labeled data is a costly task, since it demands time, resources and experts. Furthermore, some domains like disease detection suffer from unbalanced classes. These scenarios are challenging and degrade the performance of machine learning algorithms. In these cases, we can use \textit{Data Augmentation} (DA) approaches to increase the number of labeled examples in a dataset. The objective of this work is to analyze the use of Generative Adversarial Networks (GANs) as DA, which are capable of synthesizing artificial data from the original data, under an adversarial process of two neural networks. The GANs are applied in the classification of unbalanced Covid-19 radiological images. Increasing the number of images led to better accuracy for all the GANs tested, especially in the multi-label dataset, mitigating the bias for unbalanced classes.

Comparison of GANs for Covid-19 X-ray classification

This work proposes a machine learning model trained from scratch to classify and detect the presence of pneumonia from a collection of chest X-ray images. Unlike most works that use deep learning to classify whether the image is of a lung with pneumonia or not, that is, two classes to achieve a remarkable classification performance, this model uses Oriented Gradient Histogram for extra features from a provided chest X-ray image and classify it into three classes, determining whether or not a person is infected with viral or bacterial pneumonia. Despite greater complexity and use of traditional machine learning techniques, the highest accuracy achieved was 91.32% more higher than works that use deep learning approaches and seeking to solve the same degree of complexity.

Classification of chest X-ray images using Machine Learning and Histogram of Oriented Gradients

This work proposes a machine learning approach to predict the prognosis
of patients with COVID-19. To assist in this task, a descriptive analysis
and relative risk estimation were performed. In addition, the importance
of variables in the perspective of machine learning algorithms was computed
and discussed. The experiments were performed with large-scale nation-wide
dataset from Brazil. The results reveal that the model developed was able to
predict the patient’s prognosis with an AUC = 0.8382. The results also point
out that the chance of death is greater among patients over 60 years old, with
comorbidities, and symptoms such as dyspnea and Oxygen saturation (< 95%),
confirming results observed in other regions of the world.

Machine Learning for Prognosis of Patients with COVID-19: An Early Days Analysis

The COVID-19 pandemic declared in March 2020 by the World Health Organization (WHO) challenged the health system of several countries with the growing number of infected people. During the pandemic's peak in Europe, the low incidence of infection in South Korea drew the international community's attention, since not long ago that country was considered the epicenter of the pandemic outside its origin, in China. The mass testing protocol and tracing policies were pointed out as the formula for South Korean success, however, in view of the high demand and little supply of diagnostic tests for COVID-19 in the market, this strategy proved to be unfeasible to be implemented mainly in countries with large populations and with few financial resources, such as Brazil. There is also the aggravating factor regarding the effectiveness of the tests currently available, especially the rapid serology test with a high rate of false negatives. In order to offer a screening method for the application of tests, this work aims to develop a predictive model for assisting the identification of COVID-19 infection in suspected patients based on data from clinical laboratory examinations, such as blood count and urine tests. The data used comes from three sources in Sao Paulo and are hosted in the COVID-19 Data Sharing/BR Repository, a shared database of Sao Paulo Research Foundation (FAPESP). The work also proposes a comparison between balanced x imbalanced dataset and traditional x ensemble algorithms for this problem.

Premium content

Classification of anatomical landmarks in the gastrointestinal tract with endoscopy images utilizing convolutional neural networks and triplet loss

Downloads

Next from BRACIS 2021

A clustering algorithm to evaluate the attitude of Brazilian researchers regarding open access research data

Similar lecture

Quaternion-Valued Convolutional Neural Network Applied for Acute Lymphoblastic Leukemia Diagnosis