EFAI RTSUITE CT HCAP-Segmentation System

The EFAI RTSuite CT HCAP-Segmentation System is an adjunctive software tool that employs deep-learning algorithms to automatically generate initial contours for an extensive set of organs at risk from adult non-contrast CT images, encompassing the head/neck, chest, and abdomen/pelvis regions to assist in radiation therapy planning. Its contouring accuracy and consistency, validated through nonclinical and clinical studies using Dice Similarity Coefficient and Hausdorff Distance metrics against manual delineations and a comparator device, support its role in streamlining clinical workflows provided there is mandatory clinician review and editing.

The EFAI RTSuite CT HCAP-Segmentation System, also known as EFAI HCAPSeg, is a standalone software device designed to assist trained radiation oncology professionals in radiation therapy treatment planning for adult patients. Its primary function is to automatically delineate an extensive list of organs-at-risk (OARs) on non-contrast CT images covering the head and neck, chest, abdomen, and pelvis regions. The system utilizes deep-learning algorithms to process input CT images in DICOM format and generates initial OAR contours, which are output as DICOM RTSTRUCT files. These contours must then be reviewed and edited by clinicians using a DICOM-compliant treatment planning system. EFAI HCAPSeg is intended as an adjunct tool and should not replace a clinician's judgment, manual contouring, or be used for primary diagnostic interpretation or lesion detection; it also does not offer a direct user interface and operates on a specialized server within a local network. The performance of EFAI HCAPSeg was evaluated through nonclinical and clinical tests. Nonclinical testing involved a dataset of 1,846 cases from Taiwan (1,410 training, 436 testing), achieving a mean Dice Similarity Coefficient (DSC) of 0.85 for OAR contouring against ground truth. A subsequent clinical study on 157 non-contrast CT cases from the United States compared EFAI HCAPSeg against another device, yielding an overall mean DSC of 0.83 and a median 95% Hausdorff Distance (HD) of 2.23 mm. The system demonstrated non-inferiority for OARs common to both devices in the Head & Neck (DSC 0.80), Chest (DSC 0.90), and Abdomen & Pelvis (DSC 0.90) regions, and achieved a mean DSC of 0.82 for OARs unique to EFAI HCAPSeg in this comparison. Performance was shown to be consistent across various subgroups (gender, age, CT manufacturer, slice thickness), although individual OAR performance varied, with detailed DSC and HD metrics provided.

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 EFAI HCAPSeg algorithm's performance, particularly for organs with lower Dice Similarity Coefficients (DSC) or higher Hausdorff Distances (HD) like A_Carotid_L/R, Duodenum, and Testis, vary with different CT scanner models, acquisition protocols, or patient anatomies not extensively represented in the provided training and testing datasets?; What is the quantifiable impact on radiation therapy treatment planning time and inter-observer variability in OAR contouring when clinicians use the EFAI HCAPSeg as an adjunct tool compared to manual contouring alone or using other semi-automated tools?; Considering the list of 68 delineated OARs, what specific challenges did the deep learning model face with the OARs unique to HCAPSeg (e.g., LN_Pelvics, Uterus, PenileBulb) during development, and how do their performance metrics compare to established literature benchmarks for similar automated segmentation tasks?; How does the software handle CT images with significant anatomical variations (e.g., due to prior surgeries, large tumors distorting anatomy) or common imaging artifacts, and what is the recommended clinical workflow for reviewing and editing the generated contours in such complex cases to ensure safety and accuracy?; Given the 'N/A' entries for some demographic data (age, gender) and CT manufacturer details in the training/testing datasets, what methodologies were employed to mitigate potential biases and ensure the robustness and generalizability of the EFAI HCAPSeg's deep learning model across diverse global patient populations and varied imaging equipment?