Jointly Optimized and Accelerated Deep Learning–Based Breast DCE-MRI Reconstruction and Pharmacokinetic Parameter Estimation

Sammanfattning

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) provides anatomical images and pharmacokinetic (PK) parameters, such as 𝐾trans, which are essential for cancer diagnosis and treatment monitoring. Conventional workflows face significant challenges: achieving high temporal resolution requires aggressive k-space undersampling, degrading image quality and PK accuracy. Moreover, standard reconstruction and PK estimation methods, such as GRASP followed by nonlinear least-squares fitting, are computationally expensive and time-consuming, limiting clinical applicability. This thesis investigates deep learning–based approaches for joint breast DCE-MRI reconstruction and 𝐾trans estimation to improve both accuracy and computational efficiency. Two novel networks, Dual-Net and Combined-Net, are developed based on the DCE-MovieNet architecture, enabling simultaneous image reconstruction and PK parameter estimation. The proposed networks are compared against sequential pipelines (DCEMovieNet + DCE-QNet) and direct PK estimation using PK-UNet. Results show that the deep learning models achieve competitive reconstruction quality compared to GRASP. While the joint models do not outperform the sequential DCE-MovieNet + DCE-QNet pipeline in image reconstruction, Combined- Net provides the most accurate and robust tumor-specific 𝐾trans estimates. PK-UNet, despite strong global 𝐾trans accuracy, exhibits substantially higher tumor-specific errors, indicating the value of leveraging feature representations from reconstruction networks for reliable tumor assessment. Furthermore, the deep learning approaches reduce reconstruction and PK estimation times from several minutes per slice to only a few milliseconds, demonstrating strong potential for real-time clinical applications. Limitations of this work include the small dataset size and limited tumor representation, which may affect generalizability. Overall, this thesis demonstrates that deep learning provides a powerful framework for accelerating DCE-MRI reconstruction and PK analysis, with joint models enhancing tumor-specific accuracy while substantially reducing computational time.