Sparse‑view computed tomography (CT) reduces radiation by acquiring fewer projections, yet conventional sampling is designed for image reconstruction rather than downstream physical inference. This work investigates task‑aware projection selection for directly predicting hemodynamics from sparse CT angiography (CTA) data. Using computational fluid dynamics (CFD) simulations of intracranial aneurysms, we synthesize fan‑beam CTA projections and train a Fourier Neural Operator (FNO) to predict two targets: the spatially resolved time‑averaged wall shear stress (TAWSS) field and its spatial average. The FNO is coupled with optimal input selection, enabling simultaneous sampling optimization and model training. We compare fixed uniform and golden‑angle sampling with three task‑aware strategies—Gumbel‑Softmax, Soft top‑K, and an agentic AI selector—across various view‑budget levels. Results show that for the resolved TAWSS, Agentic and Soft top‑K achieve the lowest mean errors, with Agentic slightly ahead, and both vastly outperform fixed sampling under extreme sparsity. For the averaged TAWSS, optimal selection only becomes critical when acquisitions are extremely sparse. Mask analysis reveals multiple distinct subsets concentrated around informative regions. Notably, masks optimized for hemodynamic prediction perform poorly for image reconstruction, highlighting a fundamental difference between task‑optimal and reconstruction‑optimal acquisition. This study demonstrates the feasibility of designing sparse acquisitions directly around downstream physical quantities.
Review: The agent can autonomously identify projection locations most informative for hemodynamic inference, preserving prediction accuracy even with very few views, suggesting that future medical imaging may benefit from task‑driven sampling strategies rather than reconstruction‑centric ones.