In a preprint posted to arXiv on April 22, 2026 (not yet peer-reviewed), Tianyuan Wang and colleagues present a dose-aware framework that pairs penalized weighted least-squares plug-and-play (PWLS-PnP) reconstruction with a deep reinforcement learning agent. The RL policy learns to select projection angles sequentially and allocate photons adaptively according to angle-dependent noise statistics, rather than the conventional uniform-angle, equal-dose strategy. All results come from numerical simulations on synthetic phantoms; the paper reports no human subjects, no clinical sample size, and no real scanner validation.

This work goes well beyond the abstract's claim of 'improved reconstruction quality.' It sits at the inflection point where AI stops merely denoising finished images and begins directing physical data acquisition. Earlier landmark papers, such as Chen et al. (2017) in IEEE Transactions on Medical Imaging on residual CNNs for low-dose CT (trained on Mayo Clinic data from the 2016 AAPM Low Dose CT Grand Challenge), operated downstream. A 2022 MICCAI paper by Zhang et al. ('Reinforcement Learning for View Planning in Sparse-Angle CT') explored angle selection but ignored per-angle dose budgeting. Wang's integration of both under explicit photon statistics modeling closes that gap and yields clearer defect detectability when projections drop below 60 or total dose is severely limited.

What most coverage has missed is the deeper pattern: reinforcement learning's migration from games (AlphaGo, 2016) to sequential physical optimization problems in medicine. Similar RL-driven acquisition loops now appear in MRI k-space sampling and adaptive radiation therapy planning. In CT, where 70 million U.S. scans occur yearly and radiation contributes an estimated 0.4–2 % of cancers (Berrington de González, NEJM 2009), moving intelligence upstream could cut cumulative exposure more effectively than any post-processing filter.

Yet limitations are substantial. Training the RL agent requires thousands of simulated episodes; inference latency on current scanner hardware is unaddressed. Generalization across patient sizes, metal implants, or different manufacturers' beam spectra remains unproven. Regulatory pathways for adaptive AI that alter scan parameters in real time are still immature at the FDA, which distinguishes 'locked' from continuously learning algorithms.

As major vendors (Siemens Healthineers, GE Healthcare) already embed AI for static dose modulation, this preprint signals the next leap: closed-loop, information-driven scanning. If clinically validated, it could measurably lower population radiation burden while maintaining or improving diagnostic power in time-critical or pediatric settings. The acceleration of AI into medical imaging is no longer hypothetical; it is actively redesigning the scanner itself.