Model-based deep learning to restore low-dose digital breast tomosynthesis images

• Autor(a) principal: Vimieiro, Rodrigo de Barros
• Data de Publicação: 2023
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/18/18152/tde-03012024-114253/
• topic: aprendizado profundo; aprendizado profundo baseado em modelo; artificial neural networks; convolutional neural networks; deep learning; digital breast tomosynthesis; image restoration; model-based deep learning; redes neurais artificiais; redes neurais convolucionais; restauração de imagens; tomossíntese digital mamária

Digital breast tomosynthesis (DBT) and full-field digital mammography (FFDM) are the most commonly used exams for breast cancer screening. In these systems, achieving high image quality is crucial for radiologists to detect the earliest signs of breast cancer and improve the patient\'s prognosis. The radiation dose is also a concern, given that these systems employ ionizing radiation. While current systems operate within safe radiation margins, there is a growing desire to minimize radiation dose without compromising image quality. To address this challenge, image restoration techniques have emerged as valuable tools to enhance image quality from low-dose (LD) acquisitions. Traditional restoration methods rely on mathematical models that represent the underlying physics of the acquisition system. Convolutional neural networks (CNN), employing modern deep learning (DL) techniques, are capable of learning the image restoration task from data and have exhibited substantial progress in recent years. This work proposes a hybrid model-based deep learning (MBDL) framework for the restoration of DBT images acquired with reduced radiation doses, benefiting from the advantages of both fields. Specifically, our hypothesis is that the combination of known mathematical models with data-based (DB) models can improve the results of purely MB or DB approaches. First, we investigate the application of a CNN architecture to restore FFDM images, also evaluating the influence of various loss functions and diverse training strategies. Second, we introduce an MBDL approach inspired by a pipeline designed to restore LD mammographic images. We use a variance stabilization transformation (VST) and known system-related parameters to introduce priors implemented as neural network layers. Considering a Poisson-Gaussian noise model, this framework operates within a VST domain, where the noise becomes approximately Gaussian, signal-independent, and with unity variance, enhancing the stability and simplicity of the learning process. Moreover, we propose a bias-residual noise loss function to control the final noise characteristics. Three different CNN architectures were tested and resulted in better performance compared with solely DB approaches. Finally, we also propose an MBDL method to restore mammographic images corrupted with spatially correlated noise. Although further validation is necessary, preliminary results indicate that the MBDL may be suitable for this task. In conclusion, the synergy of MB methods and DB approaches has great potential to be explored within the DL domain, demonstrated by the improved results over models that do not benefit from those priors.