Computational techniques for biomedical image processing

The aim of this PhD thesis is to explore the application of computational methods to address segmentation and image generation problems for different biomedical applications and imaging techniques. The developed algorithms aim to achieve various objectives, including accelerating traditionally manual or computationally slow operations, improving results accuracy, ensuring applicability across different imaging techniques and anatomical areas, and creating transparent models for easy understanding of their functionality. The thesis investigates three main research topics: 1) Statistical mechanics-based segmentation: we propose a new method based on statistical mechanics for biomedical image segmentation. This approach conceptualizes each pixel as a particle with evolving positions and static gray levels, which interact with each other to form regions of segmentation. A key aspect of this model is the integration of a dynamic diffusion term, which quantifies stochastic variations arising during image acquisition. The Boltzmann formulation of the model is efficiently simulated using a Monte Carlo approach. An optimization strategy is proposed to fine-tune the system's internal parameters. The method is evaluated on different biomedical datasets, achieving segmentation performances in terms of Dice similarity coefficient of at least 0.91 for low-complexity segmentation tasks and at least 0.67 for high-complexity datasets. Future research should aim to enhance segmentation performance in more complex segmentation tasks. 2) COVID-19 lung lesion segmentation: we present the LungQuant system, a fully-automatic deep learning (DL) pipeline designed for segmenting and quantifying COVID-19 lung lesions in computed tomography (CT) images. This system is composed of a cascade of two U-nets, a specialized convolutional neural network architecture designed for image segmentation tasks. The LungQuant system produces as output lung and COVID-19 lesion segmentation masks, the percentage of affected lung and the corresponding CT-Severity Score (CT-SS). We trained and tested all the DL models exclusively on publicly available datasets, achieving a 90\% accuracy in CT-SS classification. We are currently developing various extensions of the study, which include technical improvements of the system, a multicenter validation and a radiomics study for clinical outcome prediction. 3) Optimized magnetic resonance fingerprinting (MRF): we propose an optimized MRF framework for generating quantitative multiparametric maps in preclinical studies. This method is composed by a DL model and a hyperparameter tuning strategy that enables the simultaneous optimization of the neural network architecture, the structure of the DL model, and the supervised learning algorithm. The system reduces the mean percentage relative error of the computed maps by a factor of at least 2 and improves the computational time by at least a factor of 37, compared to the traditional reconstruction algorithm. Furthermore, our findings demonstrate that DL method allows the use of fewer MRF images and a reduced k-space sampling percentage, making MRF examinations more efficient. Future developments of this research may involve extending the proposed system to different anatomical regions or applying it to in vivo preclinical MRF.