|
|
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.
|
|
|
Odhad obličeje z řečového signálu
Zubalík, Petr ; Mošner, Ladislav (oponent) ; Plchot, Oldřich (vedoucí práce)
Hlavním cílem této diplomové práce bylo navrhnout a implementovat systém, který bude schopný odhadnout obličej na základě řeči daného člověka. Tento problém je vyřešen pomocí systému složeného ze tří modelů konvolučních neuronových sítí. První z nich je založen na architektuře ResNet a slouží pro extrahování příznaků z hlasových nahrávek. Druhým modelem je plně konvoluční neuronová síť, která převádí tyto příznaky na styly, na základě kterých bude upravován výsledný obrázek obličeje. Získané styly jsou poté předávány na vstup generátoru StyleGAN pro vygenerování výsledného obličeje. Navržený systém je implementován v programovacím jazyce Python s využitím frameworku PyTorch. V poslední kapitole práce je rozebráno a vyhodnoceno několik důležitých experimentů prováděných v rámci ladění a testování vytvořeného systému.
|
|
|
Material Artefact Generation
Rončka, Martin ; Španěl, Michal (oponent) ; Kodym, Oldřich (vedoucí práce)
Even in the age of big data, some high-quality data sets of images with distinct artefacts are hard to obtain, either due to the source being limited or due to annotation difficulty. This applies for multiple areas such as engineering and radiology. Current state of the art methods for classification and defect detection require large well-balanced data sets to be feasible. For small data sets we face the issue of overfitting and data paucity which cause models to misclassify data in favor of over-represented classes. Part of this work deals with finding a suitable method for generating realistic images for given datasets and experiments with mitigating overfitting and data paucity by generating new images based the original dataset using Conditional Generative Adversarial Networks (CGAN) and heuristic annotation generator. Three datasets were used overall during this project. Threads dataset was used during the image generation experiments phase due to it's complexity in structure. Subsequently, two additional CGAN networks have been trained, one for Ceramics and the other for brain scans from BraTS dataset. Ceramics and BraTS was later used to evaluate the effect of generated data on training classification and segmentation networks.
|
|
|
Speech Enhancement with Cycle-Consistent Neural Networks
Karlík, Pavol ; Černocký, Jan (oponent) ; Žmolíková, Kateřina (vedoucí práce)
Deep neural networks (DNNs) have become a standard approach for solving problems of speech enhancement (SE). The training process of a neural network can be extended by using a second neural network, which learns to insert noise into a clean speech signal. Those two networks can be used in combination with each other to reconstruct clean and noisy speech samples. This thesis focuses on utilizing this technique, called cycle-consistency. Cycle-consistency improves the robustness of a network without modifying the speech-enhancing neural network, as it exposes the SE network to a much larger variety of noisy data. However, this method requires input-target training data pairs, which are not always available. We use generative adversarial networks (GANs) with cycle-consistency constraint to train the network using unpaired data. We perform a large number of experiments using both paired and unpaired training data. Our results have shown that adding cycle-consistency improves the models' performance significantly.
|
|
|
Generative Adversarial Networks Applied for Privacy Preservation in Bio-Metric-Based Authentication and Identification
Mjachky, Ľuboš ; Malinka, Kamil (oponent) ; Homoliak, Ivan (vedoucí práce)
Biometric-based authentication systems are getting broadly adopted in many areas. However, these systems do not allow participating users to influence the way their data will be used. Furthermore, the data may leak and can be misused without the users' knowledge. In this thesis, we propose a new authentication method which preserves the privacy of an individual and is based on a generative adversarial network (GAN). Concretely, we suggest using the GAN for translating images of faces to a visually private domain (e.g., flowers or shoes). Classifiers, which are used for authentication purposes, are then trained on the images from the visually private domain. Based on our experiments, the method is robust against attacks and still provides meaningful utility.
|
|
|
Rekonstrukce snímku obličeje s využitím neuronových sítí
Zubalík, Petr ; Drahanský, Martin (oponent) ; Goldmann, Tomáš (vedoucí práce)
Hlavním cílem této bakalářské práce je navrhnout a implementovat model neuronové sítě, který bude schopen rekonstruovat snímky obličeje v tak nízkém rozlišení, že na nich budou rozmazány základní části obličeje. Zadaný problém rekonstrukce obličejů je vyřešen pomocí modelů založených na konvolučních neuronových sítích. První model je postaven na architektuře ResNet, kdežto druhý staví na principech generativních kompetitivních sítí. Navržené modely jsou implementovány v programovacím jazyce Python za pomoci aplikačního programového rozhraní frameworku TensorFlow. Součástí práce je i aplikace s velmi jednoduchým grafickým uživatelským rozhraním pro snadné používání modelu. V poslední části práce je rozebráno několik experimentů provedených pro otestování úspěšnosti navržených modelů
|
|
|
Generative Adversial Network for Artificial ECG Generation
Šagát, Martin ; Ronzhina, Marina (oponent) ; Hejč, Jakub (vedoucí práce)
The work deals with the generation of ECG signals using generative adversarial networks (GAN). It examines in detail the basics of artificial neural networks and the principles of their operation. It theoretically describes the use and operation and the most common types of failures of generative adversarial networks. In this work, a general procedure of signal preprocessing suitable for GAN training was derived, which was used to compile a database. In this work, a total of 3 different GAN models were designed and implemented. The results of the models were visually displayed and analyzed in detail. Finally, the work comments on the achieved results and suggests further research direction of methods dealing with the generation of ECG signals.
|
|
|
Využití strojového učení pro generování medicínských obrazů
Hrtoňová, Valentina ; Chmelík, Jiří (oponent) ; Jakubíček, Roman (vedoucí práce)
Tato práce se zabývá využitím generativních soutěživých sítí pro generování medicínských obrazů. Nejdříve jsou popsány umělé neuronové sítě se zaměřením na konvoluční neuronové sítě a generativní soutěživé sítě. Je vypracována rešerše na využití generativních soutěživých sítí v medicíně a jsou blíže popsány vybrané publikace na téma syntézy medicínských obrazů. V programovém prostředí Python je implementován model hluboce konvoluční generativní soutěživé sítě pro generování syntetických obrazů kožních lézí a model „pix2pix“ pro tři aplikace. První aplikací modelu „pix2pix“ je generování obrazů kožních lézí, druhou je generování CT obrazů axiálních řezů páteře a poslední aplikací je generování CT obrazů sagitálních řezů páteře. Na závěr jsou prezentovány a diskutovány výsledky generování medicínských obrazů pomocí generativních soutěživých sítí.
|
|
|
Reconstruction and Enhancement of Damaged Parts of Fingerprint Images
Špila, Andrej ; Rydlo, Štěpán (oponent) ; Heidari, Mona (vedoucí práce)
This thesis deals with the problem of fingerprint image reconstruction with focus on non- recoverable regions affected by various skin diseases. A generative adversarial network with learnable convolutional gabor filter layer was trained on preprocessed dataset of real fingerprint images. The work demonstrates that the trained model can reliably repair small corrupted regions of arbitrary shapes and in case of larger holes, the global quality score of reconstructed fingerprints evaluated by MINDTCT module from NIST biometric image software is increased compared to original fingerprint. A standardized format for fingerprint images that helped stabilize the results when training generative models is proposed.
|
|
|
Generative Adversarial Networks and Applications in Bioinformatics
KOLESNICHENKO, Nikita
Generative Adversarial Networks (GAN) are currently considered a state-of-the-art method for image generation. Recently, Deep Convolutional Generative Adversarial Networks (DCGAN) yielded promising results in protein contact maps generation. The algorithm generated realistic protein structures, which were less erroneous than previously used generative methods. However, DCGAN is notorious for being hard to train due to the limitations of its loss function and complications in optimization. Wasserstein Generative Adversarial Networks (WGAN) was proposed, employing the Wasserstein loss function that stabilizes training and alleviates some of the DCGAN's training problems. In this thesis, a hyperparameter grid search for DCGAN and WGAN was conducted on the CIFAR-10 dataset. Runs with different hyperparameters were compared using Fréchet Inception Distance to determine whether WGAN is more stable than DCGAN.
|
host ::
RSS.