Koios DS

Koios DS is an AI-powered software that analyzes ultrasound images of breast lesions and thyroid nodules to provide computer-assisted diagnostic information, enhancing physicians' decision-making. By employing distinct machine learning engines for breast (using BI-RADS/U1-U5 classifications) and thyroid (using TI-RADS/ATA classifications with an 'AI Adapter'), the system aims to improve diagnostic accuracy, increase consistency among interpreting physicians, and reduce image interpretation time for cancer risk assessment.

Koios DS is an artificial intelligence (AI) and machine learning (ML) based software designed to assist trained interpreting physicians in analyzing ultrasound images of breast lesions and thyroid nodules suspicious for cancer. It functions as a web application where users select Regions of Interest (ROIs) from DICOM ultrasound images, which are then processed by distinct AI engines for breast and thyroid. For breast lesions, the software automatically classifies them based on image data into ACR BI-RADS® or European U1-U5 categories, displays a confidence level indicator, and identifies lesion shape and orientation; it also features 'Smart Click' for automated ROI generation and 'Smart Calipers' for ingesting caliper data. For thyroid nodules, Koios DS generates user-editable sonographic descriptor recommendations (Composition, Echogenicity, Shape, Margin, Echogenic Foci) and offers an optional deep-learning derived cancer risk assessment via the 'AI Adapter,' which modifies the final categorization in ACR TI-RADS™ or ATA risk stratification systems to improve diagnostic performance. Performance evaluations demonstrated notable improvements in diagnostic capabilities. The updated breast engine achieved an Area Under the Curve (AUC) of 92.9% in bench testing on 900 lesions, representing a significant 5% increase in AUC and a 24% increase in specificity compared to its predicate. Clinical study data for breast functionality indicated that readers using Koios DS experienced a 0.037 AUC improvement and showed significantly increased inter-operator agreement and reduced intra-reader variability. The thyroid engine's 'AI Adapter,' when applied to ACR TI-RADS, yielded an AUC of 79.8% in bench tests. A multi-reader, multi-case clinical trial for thyroid functionality, involving 650 cases and 15 readers, showed a statistically significant improvement in average reader AUC by +0.083 (all readers/data) and +0.074 (US readers/US data) when using Koios DS. This was accompanied by enhanced sensitivity and specificity for Fine Needle Aspiration (FNA) and follow-up recommendations, and a reduction in interpretation time.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the 'AI Adapter' for thyroid nodule analysis specifically modify the point-based or risk-shift categorizations in ACR TI-RADS and ATA RSSs, and what is the underlying mechanism for this modification?; What are the specific image features and machine learning algorithms utilized by the separate breast and thyroid AI engines to derive their respective classifications and risk assessments?; Given the significant AUC improvement with the updated breast engine (92.9% vs 88.2% for the predicate), what specific changes in the algorithm or training data contributed to this enhanced performance, particularly the 24% increase in specificity?; The document mentions 'Smart Click' for breast ROI generation; how does this automated ROI selection compare in accuracy and user workflow efficiency to manual ROI selection, and are there plans to extend this to thyroid analysis?; While the system reduces inter-operator variability, what is the range of performance improvement observed across readers with different levels of experience (e.g., radiologists vs. endocrinologists vs. surgeons as per the study participants for thyroid analysis)?