DV. Target

DV.Target is a machine learning-based software application designed to automatically delineate organs-at-risk on CT images, providing critical input for radiation therapy treatment planning. Its methodology relies on processing DICOM-compliant data through deep learning algorithms to generate OAR contours, which are then managed via an administration interface and reviewed in a separate Treatment Planning System, with performance validated through non-inferiority to existing solutions using DICE scores.

DV.Target is a standalone software application designed for trained medical professionals to assist in radiation therapy treatment planning. Its primary function is to automatically delineate organs-at-risk (OARs) on CT images. The system processes DICOM-compliant CT image data through machine learning-based algorithms to generate these OAR contours. DV.Target operates without a direct user interface for image visualization; instead, users manage image uploads, auto-contouring processes, and output files (in RTSTRUCT format) via an administration interface. It is specifically intended for contouring normal organs and not for delineating tumors or clinical target volumes, requiring integration with a Treatment Planning System (TPS) or similar software for the review, editing, and approval of all generated contours. The core science behind DV.Target involves deep learning processing for its auto-contouring capabilities, necessitating installation on a specialized server. The software can delineate a total of 49 OARs across three anatomical regions: Head & Neck, Thorax, and Abdomen & Pelvis. Non-clinical performance was evaluated through three comparison studies against a predicate device (Mirada Workflow Box) and a reference device (MIM – MRT Dosimetry). These studies utilized both public (TCIA) and in-house clinical datasets, with ground truth established by physician consensus or actual clinical contours. Using the Dice-Sørensen coefficient (DICE score) as the primary metric, DV.Target demonstrated non-inferior auto-contouring accuracy for 19 OARs shared with the predicate device and for 30 additional OARs compared to the reference device, confirming its performance capabilities.

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 performance of DV.Target's machine learning algorithms for OAR delineation vary across different CT scanner manufacturers and imaging protocols, beyond the DICOM compliance requirement?; What specific machine learning model architectures are employed in DV.Target, and how were they trained and validated to ensure generalizability across diverse patient populations and anatomical variations for the 49 specified OARs?; Given that DV.Target requires a separate TPS for contour review and editing, what is the typical quantitative and qualitative impact on dosimetrist/physicist workload and contouring efficiency compared to manual delineation or other auto-segmentation tools?; The non-clinical testing showed non-inferiority based on DICE scores. Are there data on other clinically relevant metrics, such as Hausdorff distance or mean surface distance, particularly for OARs with complex geometries or those in close proximity to target volumes?; How does DV.Target handle challenging cases, such as images with significant artifacts, post-operative anatomy, or unusual patient positioning, and what are the documented failure modes or limitations of the auto-contouring algorithms in such scenarios?