Taurus targets cheaper CT labeling by giving AI fewer, clearer decisions

Drawing detailed outlines of the aorta and its branches in CT scans is time-consuming. Those labels support aneurysm measurement, treatment planning and the training of image-segmentation models. Public datasets offer a starting point, but their annotations differ: one may distinguish named vessels, while another marks only the blood-filled channel or the aneurysm’s outer boundary.

In a preprint, University of Waterloo researchers Bing Hu, Rishabh Sharma and Jatin Mehta investigate whether a language model should tackle this work as an independent coding agent or as a reviewer inside an engineered workflow. Their main finding favors the latter approach on model cost, while highlighting how much the surrounding software determines the result.

From open-ended coding to prepared choices

The researchers built Taurus, a system that first generates labels using segmentation models and automated image-processing steps. A language model then examines rendered images, makes corrections through available tools and records whether the case needs human attention.

The comparison agents, running Claude Sonnet 5.5 or Claude Opus 5.5, instead received a scan, labeling requirements and a programming environment. They wrote and ran their own segmentation code. Unlike Taurus, they had no trained segmentation models, making this a comparison of toolkits as well as workflows.

Taurus evolved through three versions. The initial system used fixed thresholds and restricted editing tools. Reviewing its output increased the median primary score on only two of six datasets, by amounts smaller than the study’s estimated variation between repeated reviews of an individual case.

The team then examined successful coding-agent strategies and incorporated them into Taurus’s automated processing. These included measuring blood brightness before setting thresholds and tracing an aneurysm toward its outer boundary. Much of the improvement came from turning those strategies into repeatable rules, rather than asking the language model to rediscover them.

The final version, Taurus-hybrid, added fixed surface smoothing and checks for major errors. When a check identified a suspicious result, such as an aneurysm outlined at roughly half its extent, the system presented candidate corrections overlaid on the scan. The model’s role became choosing whether to accept a prepared fix.

Similar scores, lower model spending

The study used six public datasets containing 939 CT volumes. The initial automated pipeline ran across that collection, while its language-model review covered 438 sampled cases. The headline comparison of the final system used a much smaller group of 38 evaluation cases: 26 selected for poor initial performance and 12 typical cases.

Because the datasets labeled different structures, the researchers scored outputs against each dataset’s own classes. Their primary measures used Dice overlap, which ranges from zero for no overlap to one for identical labels. Medians pooled across datasets were descriptive summaries, not a single uniform measure of medical accuracy.

On the 38 cases, Taurus-hybrid’s pooled median primary score was 0.805, compared with 0.802 for the free-form Opus agent. The paired statistical test detected no difference. That supports the authors’ description of similar observed performance, but does not establish equivalence.

Median model cost was $1.22 per case for Taurus-hybrid versus $2.71 for Opus. Reported median time from scan to label was 6.4 minutes versus 16.1 minutes. However, the dollar figures excluded automated processing and engineering costs, and Taurus review costs were understated by an estimated 3–6%. They were token-price equivalents, not necessarily subscription invoices.

Why the result needs qualification

The strongest caveat concerns development. Although the 38 cases were originally held out from Taurus-tuned development, the researchers designed Taurus-hybrid after examining results and later transcripts from those cases. They explicitly describe the final comparison as not blind: it demonstrates what targeted guidance can achieve, rather than independently testing a frozen system’s generalization.

Performance also depended on failure modes and scoring conventions. On two arterial-tree scans, every Taurus version produced nearly empty labels. In another dataset, a few incorrectly labeled calcium voxels could sharply reduce the score.

Nor was the model a dependable quality inspector. Its verdict identified gross failures better than subtle ones, meaning human reviewers checking only flagged cases would miss errors. No experts adjudicated the final outputs or measured correction time.

A separate training experiment reinforces that caution: segmentation models trained on Taurus-base labels performed substantially worse than models trained on human annotations. Whether labels from the improved versions close that gap remains untested.

The practical lesson is therefore narrower than replacing expert annotation: automate reproducible operations, reserve model judgment for focused decisions, and retain expert adjudication before treating the resulting labels as reliable training data.