Koios DS for Breast

Koios DS for Breast is an AI-powered software application that assists physicians by analyzing user-selected regions of interest on breast ultrasound images, employing machine learning and computer vision to automatically classify lesions into BI-RADS-aligned risk categories and characterize their shape and orientation. The system's methodology relies on a predictive model trained on extensive case data, aiming to enhance diagnostic accuracy, improve consistency among users, and reduce variability in breast cancer assessment.

Koios DS for Breast is a software application designed as a decision support system to aid trained interpreting physicians in analyzing breast ultrasound images of patients with soft tissue breast lesions. It functions as an adjunct to diagnostic ultrasound for breast cancer, utilizing machine learning and computer vision techniques. Users select up to two Regions of Interest (ROIs) from orthogonal views of a lesion, which are then processed by the system's core engine, trained on a large database of known cases. This engine automatically classifies the lesion into four BI-RADS-aligned categories (Benign, Probably Benign, Suspicious, Probably Malignant), displays a continuous confidence level indicator, and also automatically classifies lesion shape (Oval, Round, Irregular) and orientation (Parallel to Skin, Not Parallel) according to BI-RADS descriptors. The software is a web application and can also function as an image viewer for multi-modality images, but it is not intended for primary mammography interpretation, use on mobile devices, post-surgical sites, or images with overlays like doppler or elastography. Clinical performance data from a study involving 15 readers and 900 patient cases demonstrated that Koios DS for Breast improved reader performance, indicated by a statistically significant mean Area Under the ROC Curve (AUC) shift of +0.037 when the software was used with ultrasound examination (USE + DS) compared to ultrasound alone (USE). The study also showed a significant increase in inter-operator agreement (average Kendall Tau-B improved from .5404 to .6797) and a statistically significant reduction in intra-reader variability (class switching rate decreased from 13.6% to 10.8%). Non-clinical bench testing on 900 lesions reported an AUC of 88.2% for malignancy risk classification. Furthermore, the system's accuracy in classifying BI-RADS descriptors for lesion shape and orientation was found to be statistically equivalent to that of experienced MQSA certified radiologists.

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 machine learning model handle ambiguous or poor-quality ultrasound images, and what is the reported failure rate or need for manual override in such cases within the training/validation datasets?; What specific quantitative image features (radiomics) derived by the computer vision algorithm are most heavily weighted in the BI-RADS classification and shape/orientation assessment, and how was their robustness across different ultrasound equipment validated?; The study demonstrated a reduction in inter- and intra-operator variability; what specific training is prescribed for interpreting physicians, and how long does it typically take for them to achieve proficiency and consistent integration of the tool's output?; While AUC improvement is noted, what was the observed impact on specific BI-RADS category reclassification rates (e.g., shifts from BI-RADS 3 to 2, or 4a to 3) when using Koios DS, and how did this correlate with ground truth pathology?; Given the limitations (e.g., no post-surgical sites, no doppler/elastography overlays), what are the technical challenges and future research directions for expanding the algorithm's applicability to these currently excluded but clinically relevant scenarios?