Instance Hardness in cancer image analysis: Increasing the reliability of predictions in diagnostics through meta-Learning

• Grant number: 24/09183-5
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: August 01, 2024
• End date: July 31, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: André Carlos Ponce de Leon Ferreira de Carvalho
• Grantee: Rafael Souza e Silva
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Company: Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC)
• Associated research grant: 20/09835-1 - IARA - Artificial Intelligence in the Remaking of Urban Environments, AP.PCPE

Abstract

The use of Deep Neural Networks (DNNs), such as Convolutional Neural Networks and Vision Transformers, has become very popular in the classification of medical images. Despite the success of DNNs, there are significant challenges to overcome, especially regarding the explainability and reliability of these algorithms. This project aims to enhance the understanding and reliability of classification models for cancer images by exploring the potential of Instance-Level Meta-Learning based models. The proposal involves extracting metadata that allows evaluating the complexity of examples, thus identifying observations that are difficult to classify and understanding the factors contributing to this complexity.Furthermore, once the best metrics for the domain under study are identified, a meta-model capable of providing complexity estimates about new predictions will be developed. The goal is not only to increase the reliability of DNNs, but also to identify examples that classifiers have difficulty to correctly classify. To achieve these goals, experiments and a detailed experimental analyses will be conducted, exploring different difficulty estimation metrics using representative publicly available datasets from the medical field.