28 February 2023
09:00 Master's Defense IC3 Auditorium
Theme
User-Assisted Design of a Neural Network for Targeting Brain Tumors
Student
Matheus Abrantes Cerqueira
Advisor / Teacher
Alexandre Xavier Falcao
Brief summary
Segmentation of brain tumors is a topic that deep learning has dominated in recent years, showing the best results. However, the traditional method of learning these models requires a large amount of annotated data and a high computational effort. In this sense, other methodologies are presented as an alternative, such as FLIM (Feature Learning from Image Markers), which uses the network designer's knowledge to guide learning, reducing model sizes and the need for fully annotated data. FLIM has shown good results in the creation of feature extractors, estimating convolutional filters from regions marked by the user. In this sense, we propose a feature estimation method for brain tumor images using FLIM in a multistep approach, ensuring that the first layer of the convolutional network has expressive features. Among our results, we used a small U-Net network, whose encoder is learned using FLIM and the decoder is learned by the standard backpropagation algorithm, so we were able to show that our results are within the standard deviation of large brain tumor segmentation models .
Examination Board
Headlines:
| Alexandre Xavier Falcão | IC / UNICAMP |
| Fátima de Lourdes dos Santos Nunes Marques | EACH / USP |
| Marcos Medeiros Raimundo | IC / UNICAMP |
Substitutes:
| Hélio Pedrini | IC / UNICAMP |
| Leticia Rittner | FEEC / UNICAMP |