|
|
Segmentation of multiple sclerosis lesions using deep neural networks
Sasko, Dominik ; Myška, Vojtěch (oponent) ; Kolařík, Martin (vedoucí práce)
This master thesis focused on automatic segmentation of Multiple Sclerosis (MS) lesions on MRI images. We tested the latest methods of segmentation using Deep Neural Networks and compared the approaches of weight initialization by transfer learning and self-supervised learning. The automatic segmentation of MS lesions is a very challenging task, primarily due to the high imbalance of the dataset (brain scans usually contain only a small amount of damaged tissue). Another challenge is a manual annotation of these lesions, as two different doctors can mark other parts of the brain as damaged and the Dice Coefficient of these annotations is approximately 0.86, which further underlines the complexity of this task. The possibility of simplifying the annotation process by automatization could improve the lesion load determination and might lead to better diagnostic of each individual patient. Our goal was to propose two techniques that use transfer learning to pre-train weights to later improve the performance of existing segmentation models. The theoretical part describes the division of artificial intelligence, machine learning and deep neural networks and their use in image segmentation. Afterwards, the work provides a description of Multiple Sclerosis, its types, symptoms, diagnosis and treatment. The practical part begins with data preprocessing. Firstly, brain scans were adjusted to the same resolution with the same voxel size. This was needed due to the usage of three different datasets, in which the scans had been created by devices from different manufacturers. One dataset also included the skull, therefore it was necessary to remove it by an FSL tool, leaving only the patient's brain in the scan. The preprocessed data were 3D scans (FLAIR, T1 and T2 modalities), which were cut into individual 2D slices and used as an input for the neural network with encoder-decoder architecture. The whole dataset consisted a total of 6,720 slices with a resolution of 192 x 192 pixels for training (after removing slices where the mask was empty). Loss function was Combo loss (combination of Dice Loss with modified Cross-Entropy). The first technique was to use the pre-trained weights from the ImageNet dataset on encoder in U-Net network, with and without locked encoder weights, respectively, and compare the results with random weight initialization. In this case, we used only the FLAIR modality. Transfer learning has proven to increase the metrics from approximately 0.4 to 0.6. The difference between encoder with and without locked weights was about 0.02. The second proposed technique was to use a self-supervised context encoder with Generative Adversarial Networks (GAN) to pre-train the weights. This network used all three modalities also with the empty slices (23,040 slices in total). The purpose of GAN was to recreate the brain image, which was covered by a checkerboard. Weights learned during this training were later loaded for the encoder to apply to our segmentation problem. The following experiment did not show any improvement, with a DSC value of 0.29 and 0.09, with and without a locked encoder, respectively. Such a decrease in performance might have been caused by the use of weights pre-trained on two distant problems (segmentation and self-supervised context encoder) or by difficulty of the task considering the hugely unbalanced dataset.
|
|
|
Automatic Speech Detection for VHF Channel
Nováková, Mária ; Veselý, Karel (oponent) ; Szőke, Igor (vedoucí práce)
A noisy environment in air traffic communication is an unavoidable problem. The communication between the control tower and the pilot should be the most reliable and effective. That is why voice activity detection is crucial for recognising the start of the speech segment of the communicants for automated systems. The speakers take turns providing information by pressing the push-to-talk button. To detect voice activity, various approaches are used. Even though these methods are effective, machine learning can easily outshine them. Neural networks are widely used in voice activity detection as well as in other areas. Properly trained models are efficient and adaptable. In this thesis, a solution for voice activity detection together with push-to-talk detection is proposed. Proposed models are evaluated and compared. The adaptation of the GPVAD approach is discussed and compared to the proposed models. Neural networks will have their chance to once again prove that they are suitable for any task.
|
|
|
Automatic Speech Detection for VHF Channel
Nováková, Mária ; Veselý, Karel (oponent) ; Szőke, Igor (vedoucí práce)
A noisy environment in air traffic communication is an unavoidable problem. The communication between the control tower and the pilot should be the most reliable and effective. That is why voice activity detection is crucial for recognising the start of the speech segment of the communicants for automated systems. The speakers take turns providing information by pressing the push-to-talk button. To detect voice activity, various approaches are used. Even though these methods are effective, machine learning can easily outshine them. Neural networks are widely used in voice activity detection as well as in other areas. Properly trained models are efficient and adaptable. In this thesis, a solution for voice activity detection together with push-to-talk detection is proposed. Proposed models are evaluated and compared. The adaptation of the GPVAD approach is discussed and compared to the proposed models. Neural networks will have their chance to once again prove that they are suitable for any task.
|
|
|
Segmentation of multiple sclerosis lesions using deep neural networks
Sasko, Dominik ; Myška, Vojtěch (oponent) ; Kolařík, Martin (vedoucí práce)
This master thesis focused on automatic segmentation of Multiple Sclerosis (MS) lesions on MRI images. We tested the latest methods of segmentation using Deep Neural Networks and compared the approaches of weight initialization by transfer learning and self-supervised learning. The automatic segmentation of MS lesions is a very challenging task, primarily due to the high imbalance of the dataset (brain scans usually contain only a small amount of damaged tissue). Another challenge is a manual annotation of these lesions, as two different doctors can mark other parts of the brain as damaged and the Dice Coefficient of these annotations is approximately 0.86, which further underlines the complexity of this task. The possibility of simplifying the annotation process by automatization could improve the lesion load determination and might lead to better diagnostic of each individual patient. Our goal was to propose two techniques that use transfer learning to pre-train weights to later improve the performance of existing segmentation models. The theoretical part describes the division of artificial intelligence, machine learning and deep neural networks and their use in image segmentation. Afterwards, the work provides a description of Multiple Sclerosis, its types, symptoms, diagnosis and treatment. The practical part begins with data preprocessing. Firstly, brain scans were adjusted to the same resolution with the same voxel size. This was needed due to the usage of three different datasets, in which the scans had been created by devices from different manufacturers. One dataset also included the skull, therefore it was necessary to remove it by an FSL tool, leaving only the patient's brain in the scan. The preprocessed data were 3D scans (FLAIR, T1 and T2 modalities), which were cut into individual 2D slices and used as an input for the neural network with encoder-decoder architecture. The whole dataset consisted a total of 6,720 slices with a resolution of 192 x 192 pixels for training (after removing slices where the mask was empty). Loss function was Combo loss (combination of Dice Loss with modified Cross-Entropy). The first technique was to use the pre-trained weights from the ImageNet dataset on encoder in U-Net network, with and without locked encoder weights, respectively, and compare the results with random weight initialization. In this case, we used only the FLAIR modality. Transfer learning has proven to increase the metrics from approximately 0.4 to 0.6. The difference between encoder with and without locked weights was about 0.02. The second proposed technique was to use a self-supervised context encoder with Generative Adversarial Networks (GAN) to pre-train the weights. This network used all three modalities also with the empty slices (23,040 slices in total). The purpose of GAN was to recreate the brain image, which was covered by a checkerboard. Weights learned during this training were later loaded for the encoder to apply to our segmentation problem. The following experiment did not show any improvement, with a DSC value of 0.29 and 0.09, with and without a locked encoder, respectively. Such a decrease in performance might have been caused by the use of weights pre-trained on two distant problems (segmentation and self-supervised context encoder) or by difficulty of the task considering the hugely unbalanced dataset.
|
host ::
RSS.