Limbus Contour

K241837

Limbus AI Inc. · cleared 2024-10-09 · product code QKB · Radiology

Premarket evidence — what FDA accepted

Device typesamd
source quote (p.5)
“Limbus Contour is a stand-alone software medical device.”
Algorithmneural network models based on U-Net and ResUNet architectures, trained with Adam optimization algorithm and Sørensen-Dice coefficient loss function. Backpropagation is used to adjust model parameters.
source quote (p.11)
“The architecture for the neural network models used in our device borrows its primary structure from the U-Net (Ronneberger 2015) and ResUNet (Diakogiannisa 2020). We use the Adam optimization algorithm (Kingma 2014) and the Sørensen-Dice coefficient loss function (Sørensen 1948) to train the network. The models are trained with examples of medical images and the corresponding human-generated contours for the region of interest. During training, an image is shown to the model, and the model generates a contour. The contour generated by the model is compared to the ground truth human-generated contour. The error between the generated and ground truth contours is used to adjust the parameters (weights) of the model. Backpropagation is the algorithm that uses the error to adjust the model parameters.”
Adaptive (vs locked)No
source quote (p.7)
“Locked algorithm; Deep Learning model”
PCCPNo
Cybersecurity addressedFDA source did not state this

Validation studies (1)

Bench

n=23 scans

endpoints: Sørensen-Dice Similarity Coefficient (DSC)

Reported performance (3 observations)

diceas written: “Limbus Mean DSC (Bowel_Bag)”0.93979478CI 0.92253647
source quote (p.24)
“Bowel_Bag 0.93979478 0.03659061 23 0.92253647 0.752 Passed”
diceas written: “Limbus Mean DSC (Brain)”0.992205CI 0.99078444
source quote (p.24)
“Brain 0.992205 0.00251205 16 0.99078444 0.988 Passed”
diceas written: “Limbus Mean DSC (Bladder)”0.96601238CI 0.94024138
source quote (p.23)
“Bladder 0.96601238 0.05220935 21 0.94024138 0.935 Passed”

Each value carries its own analysis unit and task — never compare or pool across devices. Source: 510(k) summary PDF.

Predicate network

Postmarket — what happened after clearance

0
recalls in product code, 24mo
0
MAUDE reports in code, 12mo
—
vs code's own 3-yr baseline
0
drift signals on this device

Recall and MAUDE counts are product-code-level (reports aren't reliably attributable to one device); a recall is shown as device-attributed only when the recall record itself lists this clearance number. Signals are descriptive observables with sources — never a judgment that the device is unsafe or drifting. Snapshot 2026-07-08.

Reimbursement — how devices like this got paid

Not yet tracked — no payment pathway indexed for this clearance (the reimbursement corpus is a growing seed set).

Applicable FDA guidance — what the submission is measured against

FDA guidance documents and guiding principles applicable to 510(k) AI/ML devices in the Radiology panel. A curated reference index, not legal or regulatory advice — each item states its own status, and a draft is never binding.

Applicability is derived from the device's FDA advisory panel and pathway — cross-cutting guidances apply to every AI/ML device; panel-specific ones are flagged. Titles, dates, and links verified against fda.gov as of July 2026.

Constat Precedent · public FDA/CMS data · descriptive decision-support, not regulatory or reimbursement advice. Share this page: constat.dev/precedent/device/K241837