Name: | Description: | Size: | Format: | |
---|---|---|---|---|
2.82 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O cancro da mama é a doença maligna mais comum entre as mulheres, ultrapassando o
cancro do pulmão com 11,7% dos casos [1]. A deteção precoce é crucial, uma vez que a doença
é frequentemente assintomática nas suas fases iniciais. Embora a Ressonância Magnética (MRI)
forneça imagens de alta qualidade essenciais para uma deteção precisa e demonstre um
desempenho superior na deteção de lesões, com uma taxa de 30% a 40% maior do que a
mamografia, a sua adoção para a deteção precoce ainda é limitada devido à complexidade e aos
custos elevados (Saeed et al., 2022). Em vez disso, a mamografia é preferida por ser
significativamente mais económica e fácil de utilizar. No entanto, as imagens de MRI continuam
a ser valiosas para a investigação e para futuros avanços, podendo tornar-se mais acessíveis e
viáveis para fins de diagnóstico ao longo do tempo. O principal objetivo deste projeto de tese de
mestrado é desenvolver algoritmos e métodos, com base em técnicas de inteligência artificial
emergentes, para melhorar a deteção (anotação) automática de lesões patológicas de cancro da
mama, nomeadamente, modelos de machine learning / deep learning (ML/DL) treinados com
imagens de ressonância magnética e os metadados associados. A anotação manual por médicos
especialistas é morosa e difícil devido ao elevado volume de imagens e às complexidades da
identificação e categorização dos tumores. Nesse sentido, é uma contribuição deste trabalho o
desenvolvimento de um método que permite na fase de pré-processamento, de forma automática,
ajustar o boundingbox das lesões anotadas. Com isto, é possível melhorar o desempenho / precisão
dos modelos de ML/DL. Ao propor um conjunto de dados refinados resultante do préprocessamento, a qualidade dos dados é melhorada para além do que foi anotado manualmente.
Os algoritmos desenvolvidos extraem características críticas das imagens médicas, melhorando a
precisão e a robustez dos modelos de aprendizagem automática. As técnicas avançadas de
extração de características foram fundamentais neste projeto, tirando partido das anotações de
especialistas para representar melhor o tecido tumoral. Isto resultou em modelos mais precisos,
alcançando uma precisão média de 84,9%, com instâncias que atingiram 90,2%. Os resultados
bem-sucedidos do projeto sugerem um avanço significativo na deteção automática do cancro da
mama, reduzindo potencialmente as taxas de mortalidade e enriquecendo o conjunto de dados
para investigação futura.
Breast cancer is the most common malignancy among women, surpassing lung cancer with 11.7% of cases [1]. Early detection is crucial, as the disease is often asymptomatic in its early stages. Although Magnetic Resonance Imaging (MRI) provides high-quality images essential for accurate detection and demonstrates superior performance in detecting lesions, with a 30% to 40% higher rate than mammography, its adoption for early detection is still limited due to complexity and high costs (Saeed et al., 2022). Instead, mammography is preferred because it is significantly cheaper and easier to use. However, MRI images remain valuable for research and future advances, and may become more accessible and viable for diagnostic purposes over time. The main objective of this master's thesis project is to develop algorithms and methods, based on emerging artificial intelligence techniques, to improve the automatic detection (annotation) of pathological breast cancer lesions, namely machine learning / deep learning (ML/DL) models trained with MRI images and the associated metadata. Manual annotation by medical specialists is time-consuming and difficult due to the high volume of images and the complexities of identifying and categorizing tumours. With this in mind, the contribution of this work is the development of a method that allows the bounding box of annotated lesions to be automatically adjusted in the pre-processing phase. This will improve the performance/accuracy of ML/DL models. By proposing a refined data set resulting from pre-processing, the quality of the data is improved beyond what was manually annotated. The algorithms developed extract critical features from medical images, improving the accuracy and robustness of machine learning models. Advanced feature extraction techniques were key in this project, taking advantage of expert annotations to better represent tumour tissue. This resulted in more accurate models, achieving an average accuracy of 84.9%, with instances reaching 90.2%. The successful results of the project suggest a significant advance in the automatic detection of breast cancer, potentially reducing mortality rates and enriching the dataset for future research.
Breast cancer is the most common malignancy among women, surpassing lung cancer with 11.7% of cases [1]. Early detection is crucial, as the disease is often asymptomatic in its early stages. Although Magnetic Resonance Imaging (MRI) provides high-quality images essential for accurate detection and demonstrates superior performance in detecting lesions, with a 30% to 40% higher rate than mammography, its adoption for early detection is still limited due to complexity and high costs (Saeed et al., 2022). Instead, mammography is preferred because it is significantly cheaper and easier to use. However, MRI images remain valuable for research and future advances, and may become more accessible and viable for diagnostic purposes over time. The main objective of this master's thesis project is to develop algorithms and methods, based on emerging artificial intelligence techniques, to improve the automatic detection (annotation) of pathological breast cancer lesions, namely machine learning / deep learning (ML/DL) models trained with MRI images and the associated metadata. Manual annotation by medical specialists is time-consuming and difficult due to the high volume of images and the complexities of identifying and categorizing tumours. With this in mind, the contribution of this work is the development of a method that allows the bounding box of annotated lesions to be automatically adjusted in the pre-processing phase. This will improve the performance/accuracy of ML/DL models. By proposing a refined data set resulting from pre-processing, the quality of the data is improved beyond what was manually annotated. The algorithms developed extract critical features from medical images, improving the accuracy and robustness of machine learning models. Advanced feature extraction techniques were key in this project, taking advantage of expert annotations to better represent tumour tissue. This resulted in more accurate models, achieving an average accuracy of 84.9%, with instances reaching 90.2%. The successful results of the project suggest a significant advance in the automatic detection of breast cancer, potentially reducing mortality rates and enriching the dataset for future research.
Description
Keywords
Cancro da mama Inteligência artificial Deteção precoce Ressonância magnética Pré-processamento Ajuste de anotações Extração de características Aprendizagem automática