| Abstract [eng] |
The application of Artificial Intelligence (AI) methods has been a successful, fast-growing, and popular domain in many fields of research for effective decision-making. However, despite the prevailing growth in AI application in medical domain, there is still a major challenge related to small or limited available datasets. Recent literature has shown that the application of data augmentation methods is a viable technique for solving small data classification problems. The increasing application of data augmentation methods in improving efficiency in computer vision tasks has continued to gain research interest also in medical imaging. Regardless of the success of data augmentation methods, there are still some negative factors affecting the performance of AI models, namely, complexity in identifying relevant features in data, generation of unrealistic synthetic images, model overfitting, and high computational complexity of model training, etc. In order to address the highlighted challenges, this dissertation introduces advanced data augmentation methods aimed at increasing size, diversity and resilience of training dataset. The choice of proposed data augmentation technique is based on the uniqueness of each analyzed dataset. Four effective augmentation method are proposed: Chebyshev orthogonal functions with probability density functions (PDFs), Voronoi Decomposition Random Region Erasing (VDRRE), Covariant SMOTE, and the combination of noise injection and colour space transformation techniques. The proposed augmentation methods are able to effectively improve classification accuracy, reduce overfitting of models and enhance model generalization. |
