Deep neural networks and its application

Ing. John Marfo

Diplomová práce

Abstract:

Abstract This thesis explores the field of deep neural networks and their diverse applications, with a focus on the classification of medical images using chest X-ray data for pneumonia detection. It presents a comprehensive evaluation of modern architectures, including Convolutional Neural Networks (CNNs), Visual Geometry Group (VGG)-based networks, and Capsule Networks (CapsNet), highlighting their strengths and limitations in a critical healthcare application. Open-source frameworks such as TensorFlow and Keras are employed to implement and experiment with these models, enabling reproducible and practical deep learning solutions. The study utilizes the publicly available Chest X-Ray Pneumonia dataset, applying multiple models to evaluate performance in terms of accuracy, precision, recall, Fβ-score, and training stability. Among the models assessed, the Hybrid CNN VGG model emerged as the best performer, offering the highest Fβ-score and a strong balance between precision and recall. Simpler models such as Vanilla CNNs performed well in recall but suffered from high false positives. Capsule Networks, particularly the basic variant, struggled with convergence, while the modified version showed some promise but remained unstable due to training irregularities like NaN losses. Through this evaluation, the thesis concludes that model architecture, data handling, and training regularization play a crucial role in the reliability of deep learning solutions in medical diagnostics. The findings not only demonstrate the practical implementation of deep learning for real-world healthcare challenges but also contribute to the discourse on model selection, optimization, and generalization in neural network design. Keywords: Convolutional Neural Network (CNN), Deep Learning, Capsule Network, TensorFlow, Kera’s, Artificial Intelligence, Hybrid VGG Model, Chest X-ray, Medical Image Diagnosis

Keywords

Capsule Network, Hybrid VGG Model, Medical Image Diagnosis, Convolutional Neural Network (CNN, Deep Learning, TensorFlow, Kera’s, Artificial Intelligence, Chest X-ray

Jazyk práce: angličtina

Datum vytvoření / odevzdání či podání práce: 3. 12. 2025

Obhajoba závěrečné práce

• Obhajoba proběhla 3. 6. 2026
• Vedoucí: doc. Ing. Arnošt Veselý, CSc.
• Oponent: Richard Vágner, externi

Česká zemědělská univerzita v Praze

Provozně ekonomická fakulta

Magisterský studijní program:
Informatics