A carregar...
Publicação
Deep Learning with scarce resources for histopathology image analysis
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.01 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Computer-Aided Diagnosis is crucial in pathology, notably with the introduction of whole
slide imaging (WSI). Deep learning excels in WSI analysis, yet challenges persist due to high
resolution (e.g., 100,000 x 100,000 pixels) and labor-intensive pixel-level annotations.
Recent advancements in Deep Learning have proven to be effective to address these issues.
More specifically, Multiple Instance Learning (MIL) and Self-Supervised Learning (SSL) are the
current state-of-the-art methods to overcome the limitations associated with WSI analysis. MIL
is ideally suited for histopathology image analysis, as it enables to learn from just WSI labels
without requiring the need for pixel-level annotation. Additionally, SSL has proven to excel at
extracting rich and meaningful representations from large unlabeled datasets which is specifically
well-suited for high resolution WSI since it doesn’t demand pixel-level annotations.
This work is build on a prior architecture that uses both methods, in which we extend the
research by incorporating other state-of-the-art SSL techniques that outperform the current SSL
method employed, such as MoCo, DINO and SwAV. We trained from scratch some of the SSL
models implemented, while the others, we used their pre-trained version due to computational
bottlenecks. Even though there is downsides regarding using pre-trained models in which were
pre-trained on different domains, some of the best results we got was from pre-trained models,
proving once more that SSL is a promising technique when dealing with lack of annotations.
Descrição
Tese de mestrado, Engenharia Informática, 2025, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Aprendizagem Auto-Supervisionada Aprendizagem de Instâncias Múltiplas Aprendizagem Profunda Images Histopatológicas Imagens de Lâmina Completa Teses de mestrado - 2025