Classification and Labeling of Medical Images based on Active Learning

• Grant number: 23/03462-7
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: May 01, 2023
• End date: August 31, 2024
• Field of knowledge: Health Sciences - Medicine - Medical Clinics
• Principal Investigator: Paulo Mazzoncini de Azevedo Marques
• Grantee: Ivar Vargas Belizario
• Host Institution: Faculdade de Medicina de Ribeirão Preto (FMRP). Universidade de São Paulo (USP). Ribeirão Preto , SP, Brazil
• Associated research grant: 16/17078-0 - Mining, indexing and visualizing Big Data in clinical decision support systems (MIVisBD), AP.TEM

Abstract

Today the complexity of the unstructured nature of large datasets is known as the term Big Data and presents several challenges in tasks involving the prediction of new data. Specifically, Big Data in medical image classification presents the challenge that deals with the existence of a high number of unlabeled images compared to a reduced number of images labeled by specialists in the process named as annotation. Furthermore, the lack of visual and specialized tools to assist with annotation tasks further increases the challenge. This type of challenge was perceived in the development of research on the image sets of the project "Mining, indexing and visualizing big data in clinical decision support systems - (MIVisBD)". The computational methods proposed to predict the label of new images, such as traditional machine learning (ML) and deep learning (DL) algorithms, need large amounts of labeled images to guarantee their efficiency. In this scenario, the annotation of large sets of medical images becomes crucial, but performing this task implies high costs in analysis time for specialists and could even be intractable in some cases. In recent years, active learning (AL), a type of semi-supervised learning, is gaining research attention for tackling the problems of both ML and DL approaches. This type of learning includes user participation in data annotation. The idea of AL algorithms is that, through query strategies, the most representative samples of the dataset are selected to be labeled by experts. Contrary to how it is done in classification by ML and DL algorithms, the objective of active learning is to reduce the effort required to label large amounts of data and, simultaneously, enable the training of models with a smaller amount of labeled data. This project proposes the development of AL-based methodologies and visualization techniques for the classification and annotation of medical images from the image sets being studied in the MIVisBD project. It is expected that the developed methodologies provide query strategies to select the most representative samples and then be used in model training, thus contributing to reducing the annotation effort of specialists by using a reduced number of labeled images.