SugarBug (1.x)
K250264Bench7, Inc. · cleared 2025-11-07 · product code MYN · Radiology
Premarket evidence — what FDA accepted
source quote (p.5)
“SugarBug is a software as a medical device (SaMD) that uses machine learning to label features that the reader should examine for evidence of decay.”
source quote (p.5)
“SugarBug uses convolutional neural network to perform a semantic segmentation task. The algorithm goes through every pixel in an image and assigns a probability value to it for the possibility that it contains decay. A threshold is used to determine which pixels are labeled in the device's output.”
source quote (p.5)
“The software reads the selected image using local processing; images are not imported or sent to a cloud server any time during routine use.”
Validation studies (2)
Standalone
n=400 images
Reader study (MRMC)
n=300 images
endpoints: The primary objective was to determine whether SugarBug improves diagnostic performance, as measured by weighted Alternative Free-response Receiver Operating Characteristic (wAFROC) area under the curve (AUC).; Secondary objectives included evaluating reader changes in sensitivity, specificity, and annotation quality (DICE scores), as well as assessing standalone model performance of SugarBug.
Reported performance (8 observations)
source quote (p.8)
“Aided readers' lesion-level mean sensitivity was 0.674 (0.615, 0.728) while that of unaided readers was 0.540 (0.445, 0.621).”
source quote (p.8)
“The mean unaided reader wAFROC-AUC was 0.659 (0.611,0.707) while the mean aided reader wAFROC-AUC was 0.725 (0.683, 0.767).”
source quote (p.7)
“SugarBug's lesion-level sensitivity and mean FPPI were 0.686 (0.655, 0.717) and 0.231 (0.111, 0.303), respectively.”
source quote (p.7)
“The DICE coefficient versus ground truth was 0.746 (0.724, 0.768).”
source quote (p.8)
“The mean unaided reader wAFROC-AUC was 0.659 (0.611,0.707) while the mean aided reader wAFROC-AUC was 0.725 (0.683, 0.767).”
source quote (p.8)
“Aided readers' lesion-level mean sensitivity was 0.674 (0.615, 0.728) while that of unaided readers was 0.540 (0.445, 0.621).”
source quote (p.8)
“Mean DICE scores (lesion annotation similarity relative to ground truth) were 0.695 (0.688, 0.702) for unaided readings and 0.740 (0.733, 0.747) for aided readings, resulting in a mean difference of 0.045 (0.035,0.055).”
source quote (p.8)
“Mean DICE scores (lesion annotation similarity relative to ground truth) were 0.695 (0.688, 0.702) for unaided readings and 0.740 (0.733, 0.747) for aided readings, resulting in a mean difference of 0.045 (0.035,0.055).”
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).