Koios DS
K242130Koios Medical, Inc. · cleared 2024-11-15 · product code POK · Radiology
Premarket evidence — what FDA accepted
source quote (p.4)
“Koios Decision Support (DS) is an artificial intelligence (AI)/machine learning (ML)-based computer-aided diagnosis (CADx) software device intended for use as an adjunct to diagnostic ultrasound examinations of lesions or nodules suspicious for breast or thyroid cancer.”
source quote (p.4)
“Koios Decision Support (DS) is an artificial intelligence (AI)/machine learning (ML)-based computer-aided diagnosis (CADx) software device intended for use as an adjunct to diagnostic ultrasound examinations of lesions or nodules suspicious for breast or thyroid cancer. The software generates a set of user-editable sonographic nodule descriptor recommendations (Composition, Echogenicity, Shape, Margin, Echogenic Foci) along with an optional, deep-learning derived cancer risk assessment of the suspected nodule from two orthogonal views. Koios DS contains two distinct AI/ML engines to characterize breast lesions and thyroid nodules. Each system uses computer vision and machine learning techniques embedded within an engine capable of reading, interpreting, analyzing, and generating findings from ultrasound data.”
Validation studies (7)
Retrospective clinical
n=900 cases
endpoints: Malignancy Risk Classifier AUC; Sensitivity; Specificity; Sensitivity to Region of Interest; Sensitivity to Transducer Frequency; Assessment of Categorical Agreement - Shape; Assessment of Categorical Agreement - Orientation; Operating Point (PLR, NLR, PPV, NPV); Data Set Drift Analysis - Malignancy Risk Classifier AUC; Data Set Drift Analysis - Categorical Output
Retrospective clinical
n=500 patients
endpoints: AUC; sensitivity; specificity
Bench
n=650 cases
endpoints: Non-inferiority Test - Sensitivity / Specificity; Non-inferiority Test - AUC; Sub-optimal ROI Test; Detection DICE Coefficient; Non-inferiority Test - Descriptor Agreement (Composition, Echogenicity, Shape, Margin, Echogenic Foci)
Bench
n=1,600 cases
endpoints: No Match Rate; Match Time; End-to-End Breast Engine Performance (AUC, Sensitivity, Specificity); End-to-End Thyroid Engine Performance (AUC, Sensitivity, Specificity); Breast Image Matching Outcomes (Successful Match, No Match, Incorrect Match, Incorrect Image); Breast Image Matching DICE Coefficient; Thyroid Image Matching Outcomes (Successful Match, No Match, Incorrect Match, Incorrect Image); Thyroid Image Matching DICE Coefficient
Bench
n=1,910 images
endpoints: Breast Freetext Identification (Breast Side, Location Type, Clock Hour, Clock Minute, CMFN, Plane); Thyroid Freetext Identification (Thyroid Side, Pole, Region, Plane); Measurement Text Identification (Measurement Description, Measurement Value, Unit of Measurement)
Reader study (MRMC)
n=900 patients
endpoints: area under the Receiver Operating Characteristic (ROC) Curve (AUC); Kendall Tau-B correlation coefficient; class switching rate
Reader study (MRMC)
n=650 cases
endpoints: Change in average AUC; Change in average Sensitivity and Specificity of FNA; Change in average Sensitivity and Specificity of Follow-up; Inter-Reader Variability; Impact on Interpretation Time
Reported performance (11 observations)
source quote (p.19)
“System performance on the 900 cases reported an AUC of 94.5%, with a Sensitivity of 0.976 [0.960, 0.992] and a Specificity of 0.632 [0.588, 0.676].”
source quote (p.19)
“System performance on the 900 cases reported an AUC of 94.5%, with a Sensitivity of 0.976 [0.960, 0.992] and a Specificity of 0.632 [0.588, 0.676].”
source quote (p.19)
“System performance on the 900 cases reported an AUC of 94.5%, with a Sensitivity of 0.976 [0.960, 0.992] and a Specificity of 0.632 [0.588, 0.676].”
source quote (p.24)
“When applied to diagnoses made using ACR TI-RADS guidelines, the Al Adapter and descriptor predictors achieved an AUC of 79.8%, demonstrating a significant increase over the average physician AUC.”
source quote (p.24)
“When recommending biopsy, the system's sensitivity is 0.644 [0.545, 0.744] and specificity is 0.612 [0.566, 0.658].”
source quote (p.24)
“When recommending biopsy, the system's sensitivity is 0.644 [0.545, 0.744] and specificity is 0.612 [0.566, 0.658].”
source quote (p.24)
“When recommending follow-up, the system's sensitivity and specificity are 0.879 [0.812, 0.946] and 0.495 [0.446, 0.544], respectively.”
source quote (p.24)
“When recommending follow-up, the system's sensitivity and specificity are 0.879 [0.812, 0.946] and 0.495 [0.446, 0.544], respectively.”
source quote (p.41)
“+0.083 [0.066, 0.099] (parametric)”
source quote (p.41)
“+ 0.084 [0.054, 0.113] (sensitivity)”
source quote (p.41)
“+ 0.140 [0.125, 0.155] (specificity)”
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
Recall and MAUDE counts are product-code-level (reports aren't reliably attributable to one device). 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).