Digital breast tomosynthesis: new strategies for optimizing acquisition geometry

Digital breast tomosynthesis (DBT) is a medical imaging modality that has been applied in breast cancer screening as it allows a more detailed analysis of breast structures compared to conventional digital mammography. Unlike the latter, in which two-dimensional (2D) images are obtained from two dif...

ver descrição completa

Detalhes bibliográficos
Autor: Costa, Arthur Chaves
Formato: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2025
País:Brasil
Recursos:Universidade de São Paulo (USP)
Repositorio:Biblioteca Digital de Teses e Dissertações da USP
Idioma:inglés
OAI Identifier:oai:teses.usp.br:tde-21052025-154500
Acesso em linha:https://www.teses.usp.br/teses/disponiveis/18/18152/tde-21052025-154500/
Access Level:acceso abierto
Palavra-chave:artefatos de reconstrução
deep neural networks
digital breast tomosynthesis
ensaios clínicos virtuais
image restoration
interpolação de quadros de vídeo
reconstruction artifacts
redes neurais profundas
restauração de imagem
tomossíntese digital mamária
video frame interpolation
virtual clinical trials
Descrição
Resumo:Digital breast tomosynthesis (DBT) is a medical imaging modality that has been applied in breast cancer screening as it allows a more detailed analysis of breast structures compared to conventional digital mammography. Unlike the latter, in which two-dimensional (2D) images are obtained from two different views of a three-dimensional (3D) volume, DBT is a technique in which multiple X-ray projections are acquired at different angles and processed for the pseudo-3D reconstruction of the breast volume, minimizing tissue overlap. The quality of DBT images is strongly influenced by acquisition parameters such as angular range and the number of projections. Higher angular sampling inside the angular span can reduce undersampling artifacts, but this comes at the cost of distributing the total radiation dose across more projections. As a result, each individual projection receives a lower dose, leading to increased image noise and potential degradation in overall image quality. To address this trade-off and enhance DBT imaging without increasing patient radiation exposure, this work explores two innovative approaches: an image restoration pipeline to compensate for the reduced per-projection dose in high-angular-sampling DBT acquisition, and a deep learning-based video frame interpolation (VFI) model that generates synthetic projections to improve angular sampling. These interpolated images are added to the existing projections set, effectively increasing angular sampling without modifying the acquisition protocol or increasing workload. To validate these approaches, virtual phantom images were generated using virtual clinical trial (VCT) software, enabling controlled testing of different acquisition parameters. The restoration approach demonstrated improvement in contrast-to-noise ratio (CNR) for simulated breast lesions, with reduced noise and fewer artifacts in the reconstructed volume compared to a standard DBT acquisition with lower angular sampling. Similarly, the interpolation approach enhanced the CNR of calcifications and improved overall reconstruction quality, reducing artifacts. These findings highlight the potential of advanced computational techniques to refine DBT imaging, potentially improving lesion detection and diagnostic accuracy without additional radiation exposure to patients.