Measure Before You Repair

Dx-Rx is an agent in which a multimodal large language model (MLLM) measures suspected artifacts before it repairs them. It receives one 3-D T1-weighted brain volume with an unknown mix of five artifacts (noise, spike, ghosting, bias field, motion), calls five diagnostic tools and four restoration tools in any order it chooses, checks each repair, can return to any earlier version, and ends with a radiology-style report that is checked against its own run log.

One test volume, before and after

Test volume t0145 with three simulated artifacts, bias field (borderline), spike (mild) and noise (moderate), viewed in ITK-SNAP: the input on the left and the volume Dx-Rx returned with GPT-6.1 Terra on the right, with the same gray window, through the sagittal, coronal and axial slices. Levels as in the paper: mild, moderate and borderline, calibrated to radiologists' grades of real scans. A selected example: Dx-Rx removed 78.9 % of this volume's artifact error; over the 138 test volumes with a repairable artifact it removed 64.5 % on average.

The agent's report for this run

The agent's own words, shortened, with version labels and tool names put in plain words, and checked against the run log. Full report, verbatim, with every tool call.