Deterministic Quantification of Strial Capillary Leakage with an Agentic Vision-Language Model

The stria vascularis maintains endolymph, whose ion homeostasis is critical for hearing and balance. Its capillaries, sealed by tight junctions, form the blood–labyrinth barrier. Strial capillary leakage is thought to dysregulate endolymph homeostasis and contribute to hearing impairment. Leakage is assayed with transcardial fluorescein isothiocyanate (FITC)-dextran 70 kDa, isolectin B4 (IB4) counter-labeling, and confocal imaging of wholemounts. The readout is qualitative, and quantification in Imaris is manual and time-consuming. We replace that step with a deterministic algorithm. The vessel compartment is the union of two detectors: a hysteresis threshold on background-flattened IB4, and multi-scale Sato vesselness on FITC. The Sato branch finds perfused capillaries as bright tubes, which diffuse dye never forms. A guard annulus outside the vessel wall rim leaves both compartments, so a lumen's halo cannot enter the extravascular compartment. The luminal reference is the 90th percentile of FITC over the vessel mask alone, so it cannot drift with the leakage. The leakage ratio R divides mean background-corrected extravascular FITC by that reference. The leakage area fraction A is the share of extravascular tissue whose background-corrected signal reaches 20% of it. A Gemini 2.5 Pro agent runs the algorithm as a tool and inspects three quality-control figures. It may re-run with a wider rim, then names the run it trusts. Reported metrics come from that run, never from model text. On synthetic phantoms, R is monotone in leak magnitude (0.026 → 0.482) and is not inverted by leak area (0.096 → 0.311 over 25–100% coverage). It is invariant to vessel density and detector gain. On 13 expert-labeled fields, group means (R, A) are (0.075, 0.054) for 4 No Leakage Controls, (0.090, 0.051) for 3 Mild Leakage, and (0.132, 0.212) for 6 Leakage/Treatment fields. R increases across the three groups; A does not. Treatment fields separate from controls with an area under the receiver operating characteristic curve of 0.833, but the one-sided Mann–Whitney p is 0.057, so no treatment effect is claimed. Two full passes returned identical values at 1.96 s per field. The pipeline reuses Otsu thresholding, Sato vesselness, and scikit-image, and replaces manual Imaris quantification with a deterministic measurement.