|
|
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.
|
|
|
Použití self-supervised learning pro rozpoznání sportovních pozic v obraze
Olekšák, Samuel ; Kocur, Viktor (oponent) ; Herout, Adam (vedoucí práce)
Táto práca demonštruje spôsob, ako minimalizovať množstvo potrebných označených trénovacích dát pri klasifikácií športových pozícií s použitím neurónovej siete trénovanej metódou contrastive self-supervised learning. Trénovanie prebieha v dvoch etapách. V prvej sa trénuje extraktor príznakov, ktorý využíva neoznačené trénovacie obrázky extrahované z nahrávok cvičení z viacerých uhlov. V druhej etape sa s využitím malého množstva označených dát trénuje jednoduchý klasifikátor napojený na extraktor príznakov. Práca pojednáva o klasifikácií v kontexte jogových póz, avšak výsledné riešenie sa dá jednoducho aplikovať aj na iné športy v prípade získania vhodnej dátovej sady. Pri návrhu riešenia je kladený dôraz na výkon výsledného modelu, aby mohol byť použiteľný v mobilných zariadeniach. Výsledný model na dátovej sade so štyrmi označenými obrázkami na každú jogovú pózu dosiahol s využitím augmentácií vstupných dát úspešnosť 76 %. Na väčšej dátovej sade s 800 označenými obrázkami na všetky pozície je úspešnosť 82 %.
|
|
|
Self-Supervised Learning for Recognition of Hand Poses in Image
Makaiová, Lucia ; Kocur, Viktor (oponent) ; Herout, Adam (vedoucí práce)
This work focuses on using self-supervised learning for the task of hand poses recognition in image. I have used contrastive method of self-supervised learning and optimized the solution iteratively, using techniques such as early stopping, triplet mining, optimization of hyperparameters or experimenting with various model architectures. The method was implemented with Pytorch framework and Tensorboard was used for data processing and visualization. I have trained the first model using a supervised method, to obtain reference values. I have successfully matched this reference result by training a self-supervised model on Handz dataset and achieving 83% accuracy. The created solution provides findings, which can be applied to similar problems, such as recognition of sport poses. The main contribution of this work is the discovery, that self-supervised methods are particularly effective when using a labeled dataset for downstream task with just a small amount of samples, which in addition have uneven distribution of samples for individual classes. Based on these findings, it is possible to create a method for self-supervised learning for recognition of sport poses or further optimize existing solution for hand poses.
|
|
|
Analýza entit v psychoterapeutických sezeních
Polok, Alexander ; Karafiát, Martin (oponent) ; Matějka, Pavel (vedoucí práce)
Tato práce se zabývá analýzou psychoterapeutických sezení v rámci výzkumného projektu DeePsy. Jejím cílem je navrhnout a vytvořit sadu příznaků modelujících průběh sezení, jež mohou odhalit na první pohled nepatrné nuance. Zmíněné příznaky jsou automaticky extrahovány ze zdrojové nahrávky s využitím hlubokých neuronových sítí. Příznaky jsou zpracovány, porovnány napříč sezeními a graficky zobrazeny, čímž vzniká dokument plnící roli zpětné vazby o sezení pro terapeuta. Tato zpětná vazba může posloužit k profesnímu růstu a kvalitnější psychoterapii v budoucnu. Bylo dosaženo relativního zlepšení detekce řečové aktivity o 37,82 %. Byl zobecněn diarizační systém VBx ke konvergenci ke dvěma mluvčím s minimálním relativním zhoršením chybovosti o 0,66 %. Byl natrénován systém pro automatické rozpoznávání řeči, jehož chybovost je o 17,06 % relativně lepší než nejlepší dostupný hybridní model. Dále byly natrénovány systémy pro klasifikaci sentimentu, typu terapeutických intervencí a detekci překrývající se řeči.
|
|
|
Self-Supervised Learning for Recognition of Sports Poses in Image
Konečný, Daniel ; Beran, Vítězslav (oponent) ; Herout, Adam (vedoucí práce)
The goal of this thesis is to recognize sports poses in image data with a self-supervised learning approach to achieve high classification accuracy even with a low number of annotated samples. Self-supervision is obtained by using images of the same scene from multiple viewpoints at identical and different times. A convolutional neural network trained with triplet loss learns embedding vectors of sports poses and a dense neural network classifies them. The proposed self-supervised model achieves classification accuracy higher by 30-40 % than a supervised model when there are only tens or ones of annotated training samples from each class. The main contributions of this thesis are a set of semi-automatic tools to prepare a dataset for the specific training process, two datasets with sets of labels for classification, and implemented models for specific self-supervised learning. The results show that self-supervised learning is a meaningful approach for solving classification problems with very few labeled samples.
|
|
|
Použití self-supervised learning pro rozpoznání sportovních pozic v obraze
Olekšák, Samuel ; Kocur, Viktor (oponent) ; Herout, Adam (vedoucí práce)
Táto práca demonštruje spôsob, ako minimalizovať množstvo potrebných označených trénovacích dát pri klasifikácií športových pozícií s použitím neurónovej siete trénovanej metódou contrastive self-supervised learning. Trénovanie prebieha v dvoch etapách. V prvej sa trénuje extraktor príznakov, ktorý využíva neoznačené trénovacie obrázky extrahované z nahrávok cvičení z viacerých uhlov. V druhej etape sa s využitím malého množstva označených dát trénuje jednoduchý klasifikátor napojený na extraktor príznakov. Práca pojednáva o klasifikácií v kontexte jogových póz, avšak výsledné riešenie sa dá jednoducho aplikovať aj na iné športy v prípade získania vhodnej dátovej sady. Pri návrhu riešenia je kladený dôraz na výkon výsledného modelu, aby mohol byť použiteľný v mobilných zariadeniach. Výsledný model na dátovej sade so štyrmi označenými obrázkami na každú jogovú pózu dosiahol s využitím augmentácií vstupných dát úspešnosť 76 %. Na väčšej dátovej sade s 800 označenými obrázkami na všetky pozície je úspešnosť 82 %.
|
|
|
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.