BU-CAD

BU-CAD is an AI-powered software tool designed to enhance diagnostic accuracy and efficiency for breast cancer detection by assisting physicians in the analysis of breast ultrasound images, through automated identification of suspicious lesions and characterization using quantitative scores and BI-RADS descriptors. Its methodology relies on distinct modules for image viewing, lesion identification, and lesion analysis, with clinical studies demonstrating improved reader performance in AUC, sensitivity, specificity, and reduced interpretation times when using the software.

BU-CAD is a software application designed to assist trained physicians in analyzing breast ultrasound images for patients with soft tissue breast lesions suspicious for cancer. The system comprises three main components: a Viewer for loading and managing multi-modality images (ultrasound and mammography), a Lesion Identification Module that automatically identifies regions of interest (ROIs) and lesion contours on ultrasound images, and a Lesion Analysis Module. This analysis module processes user-selected or software-identified ROIs from up to two orthogonal views to generate a score of lesion characteristics (SLC) indicating malignancy likelihood (ranging from 0-100, corresponding to BI-RADS 2-5), a corresponding BI-RADS category (2, 3, 4a/b/c, 5), and BI-RADS descriptors for lesion shape, orientation, margin, echo pattern, and posterior features. The software also supports image adjustment, measurement, documentation, and structured reporting, but is not intended for primary mammography interpretation or use on mobile devices. Clinical performance evaluations demonstrated BU-CAD's effectiveness in aiding diagnosis. A multi-reader multi-case (MRMC) study involving 628 cases and 16 readers (radiologists and breast surgeons) showed that physician performance, measured by the Area Under the Curve (AUC) of location-specific ROC, significantly improved from 0.7786 to 0.8160 when aided by BU-CAD (mean AUC shift +0.0374), alongside increases in sensitivity, specificity, and a notable reduction in interpretation time by approximately 40%. A standalone study on 1139 cases (from North America, Europe, and Taiwan) reported a standalone AUC_LROC of 0.8203 for BU-CAD, with its Lesion Identification Module (CADe) achieving 93.24% accuracy in identifying lesions. The Lesion Analysis Module (CADx) also demonstrated robustness, maintaining an AUC above 0.8 even when ROIs were systematically shrunk by up to 16%. However, BU-CAD has limitations, including non-use on post-surgical excision sites or images with Doppler/elastography overlays, and patient management decisions should not solely rely on its analysis.

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 BU-CAD algorithm differentiate between various BI-RADS descriptors (e.g., shape, margin, echo pattern) based on specific ultrasound image features, and what is its standalone accuracy for each individual descriptor compared to expert radiologist consensus?; What are the specific performance metrics (sensitivity, specificity, AUC) of the Lesion Identification Module when detecting very small lesions (e.g., <1 cm), particularly those with subtle or atypical sonographic features?; The study indicates a ~40% reduction in interpretation time; were potential long-term effects, such as reader fatigue, over-reliance on the software, or changes in unaided diagnostic skill, investigated?; How does the standalone AUC_LROC of 0.8203 for BU-CAD on the 1139-case dataset compare directly to the performance of experienced radiologists specializing in breast ultrasound interpreting the same dataset without AI assistance?; Regarding the limitation for use on sites of post-surgical excision, what specific image features, artifacts, or architectural distortions commonly found in post-surgical breast tissue interfere with the software's current analytical capabilities?