Qua, 02 outubro 2024 (Todo dia)
Tipo de evento:
Seminário
Organizador:
Departamento de Fisica
Local do evento:
Sala Manuel Laranjeira Edif I
Seminário DF
Departamento de Física, Faculdade de Ciências e Tecnologia, Universidade NOVA de Lisboa, Campus da Caparica, Portugal
Anfiteatro Manuel Laranjeira - Edifício I
2 de Outubro de 2024 às 14h
An Introduction to Deep Learning
Mestre Mariana Dias
Departamento de Física, Faculdade de Ciências e Tecnologia, Universidade NOVA de Lisboa, Campus da Caparica, Portugal
Abstract:
Deep Learning has become a key driver of advancements in various fields, from medical diagnostics to autonomous systems. Its effectiveness in addressing complex problems stems from the foundational principles of neural networks, which enable algorithms to learn and adapt to intricate patterns in data. By examining the underlying mechanisms and architectures of these networks, we can better understand the capabilities and limitations of Deep Learning. In this seminar, the foundational knowledge of Deep Learning will be explored, offering both technical insights and a broader exploration of these methods.
Mariana Dias Bio:
Mariana Dias received her master’s degree in biomedical engineering from Instituto Superior Técnico, University of Lisbon, in 2020. She is a PhD student from the MIT Portugal Program on the areas of "Data Science" and "Digital Transformation in Manufacturing" and she is currently working at LibPhys-Biosignals, NOVA FCT. Her research work is focused on the development of Deep Neural Networks to process and extract patterns from biosignals, with application on the evaluation of the occupational risk from physiological data.
Mestre Helena Pereira
Departamento de Física, Faculdade de Ciências e Tecnologia, Universidade NOVA de Lisboa, Campus da Caparica, Portugal
Abstract:
The loss of neurons in the substantia nigra (SN) is a pathological feature of Parkinson's disease (PD), but its segmentation in medical imaging is challenging. This study proposes an automatic method to segment Non-PD and PD patients' left and right SN in susceptibility-weighted images (SWI) and Magnitude SWI-derived images. The images used in this study were obtained from the NTUA Parkinson's disease database: 11 SWI image volumes (4 Non-PD and 7 PD) and 8 Magnitude image volumes (3 Non-PD and 5 PD). One Non-PD template was created per each image type to obtain a mask of the left and right SN. All images were registered to the respective template. The segmentation method consisted of setting a seed point based on the masks of 2D image slices created to perform a flood fill algorithm. The method’s performance was evaluated using the dice similarity coefficient (DSC) between automatic and manual segmentation. The segmentation method achieved higher performance using the SWI images (DSC: SN = 0.684 ± 0.182; left SN = 0.722 ± 0.137; right SN: 0.664 ± 0.226) than using the Magnitude images (DSC:
SN = 0.589 ± 0.249; left SN = 0.604 ± 0.223; right SN = 0.569 ± 0.284), suggesting that the SWI images are more suitable for SN segmentation. Although not yet as effective as manual segmentation, the proposed method may aid in SN segmentation.
Helena Rico Pereira Bio:
Helena Rico Pereira has master’s degree in biomedical engineering from FCT NOVA. In 2019, Helena had a fellowship from Instituto de Biofísica e Engenharia Biomédica Universidade de Lisboa, wherein she was studying imaging biomarkers such as brain connectivity measures to better understand how neurodegenerative disorders affect the brain until 2021. Then, she got a fellowship from Fundação Ciência e a Tecnologia to pursue her PhD focused on dementia in Biomedical Engineering at FCT NOVA.
