A six-step agent turned shape statistics into named signs
The study's target was a translation problem rather than a prediction problem. Radiomics yields numbers such as sphericity, surface area, and elongation, while clinicians reason with visual descriptors such as lobulation or a cauliflower-like contour. The authors built a pipeline intended to cross that gap automatically, converting quantitative morphological features into semantic signs — named, visually intuitive descriptors that a radiologist could apply at the scanner.
The pipeline ran on the Coze platform with the DeepSeek V4 Pro model and was organized as six sequential steps. It first computed cohort-level descriptive statistics for all 14 morphological features. It then spawned five parallel analyst agents, each receiving identical prompts and identical input data, and each proposing up to three candidate signs with everyday-object names, morphological characterizations tied to quantitative values, and clinical implications. A third step pooled the candidates, merged equivalent names, and ranked each unique sign by how many of the five analysts independently proposed it. Later steps produced clinician-ready operational definitions, a sign–feature correspondence table, and automated figures.
The consensus count is the pipeline's own selection mechanism, not a measure of diagnostic accuracy. A sign proposed by four of five analysts is more reproducible within this workflow, but that frequency says nothing about whether the sign separates diseases. The paper treats all generated criteria as hypotheses awaiting external validation, and only one sign was carried forward.
Fourteen shape features proved stable before any sign was named
Tumor segmentation was performed by one radiologist with more than 10 years of experience using ITK-SNAP, tracing the enhanced portion of the tumor including internal necrosis slice by slice on contrast-enhanced T1-weighted images, with T1 and T2 images as reference. PyRadiomics then extracted 14 morphology-related features, including elongation, flatness, major and minor axis lengths, maximum 3D and 2D diameters, mesh volume, sphericity, surface area, surface-to-volume ratio, and voxel volume.
Feature stability was checked by having a second radiologist with more than 10 years of experience re-segment 30 randomly selected development cases and repeat extraction. Intraclass correlation coefficients ranged from 0.910 to 0.996, which the authors describe as excellent stability. This matters because the entire downstream reasoning chain consumes these 14 numbers; unstable inputs would make any named sign an artifact of segmentation rather than of tumor shape.
The scope here is narrow by design. Only morphological features entered the pipeline, a choice the authors justify by noting that many conventional imaging signs are morphological. Texture and wavelet features, and subregional analyses such as solid versus necrotic compartments, were left to future work.
Cauliflower and eggplant each drew four of five analysts
Across the five analyst sessions, eight distinct signs emerged. Aggregation produced three leading signs: the cauliflower sign, proposed as cauliflower in four of five sessions and as blackberry in one; the eggplant sign, proposed in four of five sessions; and the pancake sign, proposed as pancake in two sessions and as stingray in one. The merging rule matters for reading these counts — a blackberry proposal was folded into cauliflower because the aggregator judged the names morphologically equivalent, so the 4/5 figure reflects consolidation rather than five independent identical labels.
For the cauliflower sign, the pipeline described a generally rounded but distinctly bumpy outline with an irregular, knobby, deeply clefted margin. Its proposed criterion was at least three nodules with both protrusion depth and cleft depth of 5 mm or more, present on at least two adjacent axial slices on contrast-enhanced T1-weighted images, assessed by sweeping inferiorly and selecting the slice with the highest nodule count.
The sign–feature table shows how the model justified its names. Surface-to-volume ratio was cited by all five analysts who proposed the cauliflower sign, and surface area by four of five. For the eggplant sign, major axis length was cited by all four proponents; for the pancake sign, flatness by all three. These are model-generated associations between a name and a feature, not tested biological relationships.
Radiologists replaced single-slice counting with a 3D rule
Two radiologists with more than 10 years of experience reviewed the full pipeline output and agreed that nodular protrusions were the essential character of the cauliflower sign. They also identified a practical flaw: judging on a single slice could be confounded by slice thickness and slice angulation. The study used 6 mm slice thickness, so a single axial image is a coarse sample of a three-dimensional surface.
The modified criterion therefore required at least three protrusions counted cumulatively across all slices, each with base width of at least 5 mm and projection depth of at least 5 mm, located outside the main tumor contour, and visible on at least two orthogonal planes. This is a substantive human edit, not a cosmetic one: the AI's slice-selection instruction was discarded and the measurement basis changed from one slice to the whole volume.
Any claim that the pipeline produced a validated sign should be read against this step. What was validated externally was the clinician-modified cauliflower criterion, applied to contrast-enhanced T1-weighted images. The eggplant and pancake signs, and the AI's original single-slice rule, were never tested on independent data.
The modified sign separated GBM from metastasis with AUC 0.75
Validation used an external cohort of 50 glioblastomas and 50 solitary brain metastases, with cases presented in random order and both radiologists blinded to pathology. The two readers reported the sign in a mean of 73% of glioblastomas (66% and 80%) and a mean of 23% of metastases (20% and 26%). Both readers' comparisons between the two tumor types were significant at p < 0.001, and inter-rater agreement was substantial with a kappa of 0.76.
The resulting AUCs were 0.73 (95% CI 0.64–0.82) for the first radiologist and 0.77 (95% CI 0.69–0.85) for the second, mean 0.75, with no significant difference between readers (p = 0.229). The operating characteristics differ in a way that matters clinically. Radiologist 1 had sensitivity 66.0%, specificity 80.0%, accuracy 73.0%, positive predictive value 76.7%, negative predictive value 70.2%, positive likelihood ratio 3.30, and negative likelihood ratio 0.42. Radiologist 2 had sensitivity 80.0%, specificity 74.0%, accuracy 77.0%, positive predictive value 75.5%, negative predictive value 78.7%, positive likelihood ratio 3.08, and negative likelihood ratio 0.27.
The two readers therefore traded sensitivity against specificity while landing at similar AUCs. A negative likelihood ratio of 0.42 versus 0.27 is not a trivial difference if the sign were used to argue against glioblastoma, which is why the mean AUC alone understates how the readers behaved. These figures also come from a single external cohort of 100 patients at one hospital, so the confidence intervals, not just the point estimates, are the honest summary.
Named signs target a different gap than saliency maps
The authors position their approach against two alternatives. Conventional radiomics workflows produce numerical feature groupings that require post-hoc expert labeling, and explainability techniques such as saliency maps, permutation importance, and Shapley values describe how a model uses features rather than giving clinicians a descriptor they can apply by eye. The paper's claim is that a semantic sign sits closer to routine visual reading, and can therefore serve as a form of clinical translation for radiomics.
The proposed division of labor is explicit: the AI generates hypotheses, and the clinician validates, contextualizes, and decides. The authors attribute the model's ability to name signs from numbers alone to morphological knowledge absorbed during pretraining on medical literature and image-caption pairs, which they argue implicitly links shape descriptors to visual phenotypes. That explanation is the authors' interpretation of an observed capability, not something the study tested.
This is also where the evidence is thinnest relative to the framing. The pipeline's sign–feature associations are model-generated hypotheses; the paper states that only the cauliflower sign was validated, as a proof of concept. A reader comparing this to radiomics explainability work should note that the comparison is conceptual, since no head-to-head comparison with saliency or Shapley methods was run.
One sign, one center, one model, one tumor type
The pipeline was evaluated on a single-center, single-tumor-type dataset with a limited sample size, using MRI from multiple scanners with relatively thick 6 mm slices, and with only one AI platform and one language model. Stability and generalizability across tumor types, imaging protocols, and model architectures remain untested, so the 4/5 and 5/5 consensus counts should be read as reproducibility within this configuration rather than as a property of agentic AI in general.
Two further boundaries shape how far the AUC result travels. The signs are inherently three-dimensional, while radiologists read two-dimensional images, so applying them requires a mental mapping the paper acknowledges as a current burden. And no prospective study has evaluated whether the cauliflower sign changes radiologist workflow, diagnostic confidence, or downstream management; the reported performance is diagnostic discrimination on retrospective images, not care delivery.
The practical priority that follows is specific: the value demonstrated here is the generation of a testable, visually meaningful criterion from shape features, and the one criterion that survived external testing needed human revision before it worked. The next decisive test is whether the unmodified pipeline output, or the untested eggplant and pancake signs, hold up under multicenter validation — not whether the framework can produce more signs.
Sources used in this answer
From quantitative features to imaging signs: agentic AI-driven autonomous discovery in glioblastoma
An agentic AI pipeline converted 14 morphological radiomic features from 106 glioblastomas into eight candidate signs, and the clinician-modified cauliflower sign reached AUCs of 0.73 and 0.77 across two blinded radiologists separating 50 GBMs from 50 metastases.