TRANSFER OF LEARNING TO RECOGNITION OF MEDICAL IMAGES RELATED TO CANCER.

• Grant number: 24/03429-2
• Support Opportunities: Scholarships in Brazil - Doctorate
• Effective date (Start): August 01, 2024
• Effective date (End): August 31, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computational Mathematics
• Principal Investigator: André Carlos Ponce de Leon Ferreira de Carvalho
• Grantee: Bruna Christina Battissacco
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated research grant: 21/11905-0 - Center of Science, Technology and Development for innovation in Medicine and Health: inLab.iNova, AP.CCD

Abstract

Cancer, one of the main causes of death in the year 2020, is a heterogeneous disease with several subtypes. Worldwide estimates, with a projection for 2040, point to an increase in the number of cases of the disease, highlighting that there are cancers that have a higher incidence in the general population, including breast, lung, prostate and colorectal cancer. Thus, it is necessary to prepare for the impact of cancer on the health system by adopting preventive measures, early diagnosis and allocation of resources for adequate treatment. In this context, for a better diagnosis and prognosis of cancer, patients commonly undergo imaging tests. Reliable imaging is critical for assessing the degree of cancer spread and post-treatment surveillance. Medical images with the aid of Artificial Intelligence provide physicians with information for faster and more accurate reports. However, to induce models with good predictive performance, such as neural networks, a training set with a high number of labeled examples is commonly required. For this reason and for having to adjust parameters of several layers, the training algorithms used for the adjustment of these networks have a high computational cost. One of the main alternatives to reduce the need for a large training set and a long training period is the transfer, to the network to be trained, of knowledge acquired by another network in a similar data set, a process called Learning by Transfer. The objective of this project is to evaluate the benefits of methods based on transfer of learning for recognizing medical images related to cancer. For this, specific alternatives will be investigated, which include the development of methods based on Transfer Learning, also associated with Convolutional Neural Networks, Vision Transformer and considering the new state-of-the-art Artificial Intelligence architecture called Pathways, and other subfields of Learning of Machine. Challenges addressed include limitation of the number of medical images, low number of labeled images, and the need to generate predictive models for diagnosing cancer variations. The present work proposal presents great potential for collaboration and contribution to advances in innovative Artificial Intelligence tools by seeking to overcome restrictions related to obtaining medical images and achieving better predictive performance than through the use of generic learning transfer strategies.