Self-supervised learning guided by chaos-based transforms in aiding the diagnosis of Lung Cancer

• Grant number: 24/01245-1
• Support Opportunities: Regular Research Grants
• Start date: May 01, 2024
• End date: April 30, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Joao Batista Florindo
• Grantee: Joao Batista Florindo
• Host Institution: Instituto de Matemática, Estatística e Computação Científica (IMECC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated researchers: Gwanggil Jeon ; Konradin Metze

Abstract

The increasing availability of digitized data and the success of deep learning algorithms have led to these models being increasingly used in medical image analysis. One difficulty, however, that frequently arises in this type of application is the difficulty in obtaining a large amount of annotated data. This has recently led to a growing interest in the self-supervised learning approach, in which pre-training of the network can be done on a set of images that do not require annotation. On the other hand, these algorithms usually explore different views of the original image and can benefit from more elaborate transformations applied to these images. In this context, this project aims to present a hybrid image analysis approach that combines partially self-supervised visual \textit{transformers} and chaos-based transforms. The methodology will be applied to aid in the diagnosis of lung carcinoma using cytology and histology images. Lung cancer is one of the deadliest and most common types of cancer worldwide, being the one that kills the most among men and the second most among women. It is expected that this project will facilitate the development of a highly efficient computational system that will immediately benefit a tool to assist specialist doctors and allow for a faster and more accurate diagnosis. The system will also allow the interpretability of results associated with the model's attention maps. In addition to this specific application, the system can be easily adapted to other related contexts, so that benefits will be obtained for the area of digital pathology and even computer vision as a whole. (AU)