Publicação
Automatic segmentation and classification of brain tumors based on multisequence MRI images with deep learning methods
| Resumo: | Gliomas are the most common primary brain tumors. Unfortunately, these neoplasms hold the worst prognosis among all brain tumors, as well. They can be broadly categorized as low or high grade gliomas. Magnetic Resonance Imaging is the standard imaging technique for their assessment. Using it, physicians can extract measurements that are crucial for treatment planning and follow-up. Notwithstanding, manual segmentation is time-demanding and prone to variability. Also, tumor grading by biopsy is very important, but it is invasive, and prone to sampling error. Therefore, automatic approaches for both segmentation and grading are needed. However, these tasks are quite challenging due the the heterogeneity of gliomas, as well as the variability among Magnetic Resonance Imaging scans. This makes it difficult to model brain tumors from prior knowledge. Machine Learning algorithms can learn how to perform a task directly from the data. Some of these algorithms may be categorized as Representation Learning if they can learn features directly from the data. Among these methods, Deep Learning is a group of Representation Learning algorithms that learn multiple levels of representations. In the past years, Deep Learning-based methods have shown remarkable performances. Hence, the aim of this work was to investigate Deep Learning methods, and use them for the automatic segmentation and grade classification of brain tumors in multisequence structural Magnetic Resonance Imaging. Additionally, an often cited setback of these complex models is their “black box” behavior. Thus, in this work we also studied interpretability of Machine Learning algorithms applied to medical imaging. Therefore, we built our work on: brain tumor image analysis in Magnetic Resonance Imaging, Machine Learning with focus on Representation Learning, and its interpretability. We investigated Convolutional Neural Networks for the task of segmentation. As a starting point, we studied a classification Convolutional Neural Network. We were able to show its effectiveness, as well as the importance of careful pre-processing. However, afterwards we adopted a more efficient Fully Convolutional Network approach. In this setting, we proposed a hierarchical approach for dealing with class imbalance. Finally, the relationships among channels of feature maps were studied. We proposed and showed the benefits of recombination and recalibration of feature maps in the context of Fully Convolutional Networks for semantic segmentation. Automatic glioma grading from structural Magnetic Resonance Imaging images is challenging due to their large heterogeneity. Additionally, a tumor mass must be graded as a whole. Therefore, we propose 3D Convolutional Neural Networks for automatic glioma grading. Since Convolutional Neural Networks learn features directly from the data, it allows one to bypass the need for a very accurate segmentation that is often seen in radiomics-based approaches, which use hand-crafted features. “Black box” systems may pose trusting issues when deployed in critical domains, such as the medical field. This is due to professionals not being able to explain certain predictions. Therefore, interpretability of machine learning systems is a crucial field of research, given the high performances currently achieved with these systems. We first investigated this topic in a Restricted Boltzmann Machine and Random Forest classifier system in the context of segmentation. We proposed methodologies for both global and local interpretability. The former is targeted at understanding if the system learned the relevant relations in the data, while the latter is focused on explaining individual predictions. We were able to confirm if the system learned correct patterns, but we also found a bias in the database. Later, we employ interpretability methodologies to inspect the 3D Convolutional Neural Network for glioma grading. With it, we were able to catch and correct an issue during pre-processing. Hence, we provide tools and study cases that show how interpretability not only helps in increasing trust, but it may also be useful during the development cycle. Finally, all methodologies developed in this work were validated in publicly available databases. This ensures a fair comparison with the state of the art. Additionally, it enables future work to be directly compared with us. |
|---|---|
| Autores principais: | Pereira, Sérgio Rafael Mano |
| Assunto: | Ciências Naturais::Ciências da Computação e da Informação |
| Ano: | 2019 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Gliomas are the most common primary brain tumors. Unfortunately, these neoplasms hold the worst prognosis among all brain tumors, as well. They can be broadly categorized as low or high grade gliomas. Magnetic Resonance Imaging is the standard imaging technique for their assessment. Using it, physicians can extract measurements that are crucial for treatment planning and follow-up. Notwithstanding, manual segmentation is time-demanding and prone to variability. Also, tumor grading by biopsy is very important, but it is invasive, and prone to sampling error. Therefore, automatic approaches for both segmentation and grading are needed. However, these tasks are quite challenging due the the heterogeneity of gliomas, as well as the variability among Magnetic Resonance Imaging scans. This makes it difficult to model brain tumors from prior knowledge. Machine Learning algorithms can learn how to perform a task directly from the data. Some of these algorithms may be categorized as Representation Learning if they can learn features directly from the data. Among these methods, Deep Learning is a group of Representation Learning algorithms that learn multiple levels of representations. In the past years, Deep Learning-based methods have shown remarkable performances. Hence, the aim of this work was to investigate Deep Learning methods, and use them for the automatic segmentation and grade classification of brain tumors in multisequence structural Magnetic Resonance Imaging. Additionally, an often cited setback of these complex models is their “black box” behavior. Thus, in this work we also studied interpretability of Machine Learning algorithms applied to medical imaging. Therefore, we built our work on: brain tumor image analysis in Magnetic Resonance Imaging, Machine Learning with focus on Representation Learning, and its interpretability. We investigated Convolutional Neural Networks for the task of segmentation. As a starting point, we studied a classification Convolutional Neural Network. We were able to show its effectiveness, as well as the importance of careful pre-processing. However, afterwards we adopted a more efficient Fully Convolutional Network approach. In this setting, we proposed a hierarchical approach for dealing with class imbalance. Finally, the relationships among channels of feature maps were studied. We proposed and showed the benefits of recombination and recalibration of feature maps in the context of Fully Convolutional Networks for semantic segmentation. Automatic glioma grading from structural Magnetic Resonance Imaging images is challenging due to their large heterogeneity. Additionally, a tumor mass must be graded as a whole. Therefore, we propose 3D Convolutional Neural Networks for automatic glioma grading. Since Convolutional Neural Networks learn features directly from the data, it allows one to bypass the need for a very accurate segmentation that is often seen in radiomics-based approaches, which use hand-crafted features. “Black box” systems may pose trusting issues when deployed in critical domains, such as the medical field. This is due to professionals not being able to explain certain predictions. Therefore, interpretability of machine learning systems is a crucial field of research, given the high performances currently achieved with these systems. We first investigated this topic in a Restricted Boltzmann Machine and Random Forest classifier system in the context of segmentation. We proposed methodologies for both global and local interpretability. The former is targeted at understanding if the system learned the relevant relations in the data, while the latter is focused on explaining individual predictions. We were able to confirm if the system learned correct patterns, but we also found a bias in the database. Later, we employ interpretability methodologies to inspect the 3D Convolutional Neural Network for glioma grading. With it, we were able to catch and correct an issue during pre-processing. Hence, we provide tools and study cases that show how interpretability not only helps in increasing trust, but it may also be useful during the development cycle. Finally, all methodologies developed in this work were validated in publicly available databases. This ensures a fair comparison with the state of the art. Additionally, it enables future work to be directly compared with us. |
|---|