Author: Mark Anastasio, Hua Li, Zhuchen Shao 👨🔬
Affiliation: Washington University School of Medicine, University of Illinois Urbana-Champaign 🌍
Purpose: Ill-conditioned reconstruction problems in medical imaging, such as those arising from undersampled k-space data in MRI, can result in degraded image quality and clinical task-orientated performance. While objective measures of image quality (IQ) have been widely employed in the evaluation and refinement of image reconstruction methods, there are relatively few studies in which such IQ measures are explicitly incorporated into the design of a reconstruction method. This study investigates the incorporation of a data consistency (DC) mechanism into a learned task-informed reconstruction method. The DC mechanism reduces reconstruction error by enforcing physical constraints, while task-relevant information is preserved through the explicit integration of objective IQ measures.
Methods: A task-informed reconstruction framework is designed, integrating DC mechanisms within a U-Net architecture. The framework is optimized using a hybrid loss function that combines a traditional reconstruction and task-specific terms, enhancing clinically relevant features. The DC mechanism enforces consistency with the raw k-space measurements. Extensive experiments were conducted on undersampled MRI reconstruction datasets with binary signal detection tasks. The framework was tested under real-world conditions involving shifts in observers or tasks between training and inference, providing insights into its robustness.
Results: The task-informed DC U-Net demonstrated superior performance over the standard U-Net by reducing hallucinations and improving reconstruction quality. Increasing the task-based loss weight in the hybrid loss enhanced signal detectability. In observer shift, a higher task-based loss mitigated its negative effects. For task shift, training on complex tasks with broader amplitude ranges reduced the performance drop caused by the shift.
Conclusion: The integration of the DC mechanism with objective IQ measures in a deep learning framework significantly improved reconstruction quality and task-relevant information related to signal detection tasks. This study provides insights into clinically relevant scenarios, including observer and task shifts, offering guidance for the design of robust medical imaging methods.