Auto Segmentation

Auto Segmentation is a software tool designed to accelerate radiation therapy planning by automatically generating initial organ-at-risk contours from CT images using deep learning algorithms. The system processes DICOM images from adult patients and outputs a Radiotherapy Structure Set for review and refinement by clinicians, with its performance validated through quantitative metrics like the DICE similarity coefficient and qualitative clinical evaluations.

Auto Segmentation is a post-processing software tool developed by GE Medical Systems, LLC, designed to aid in radiation therapy planning for adult patients. Its primary purpose is to accelerate the clinical workflow by automatically generating initial contours of organs at risk (OARs) from Computed Tomography (CT) images, which can then be reviewed and modified by dosimetrists, medical physicists, and radiation oncologists. The system utilizes deep learning algorithms to segment up to 40 OARs or organ subregions across the head and neck, thorax, abdomen, and pelvis. It processes both contrast and non-contrast CT images acquired from various scanners and outputs the segmented OARs as a DICOM Radiotherapy Structure Set (RTSS), which is automatically networked to review workstations for user verification and correction as needed. The underlying technology relies on deep learning models trained on a dataset of 911 CT exams and subsequently validated on a separate, independent dataset of 302 retrospective CT radiation therapy planning exams from diverse global sources. Performance was quantitatively assessed using the DICE similarity coefficient (DSC) by comparing the software-generated contours against ground truth annotations established by qualified radiotherapy practitioners. The results indicated that Auto Segmentation's performance, with mean DSC values such as 92.53% for Left Kidney and 81.69% for Esophagus, is in line with or superior to existing automated contouring devices. A qualitative preference study further confirmed the adequacy of the generated contours for radiotherapy planning use, and subgroup analyses showed consistent algorithm performance across different CT systems, patient demographics, and image characteristics, with known limitations described in user documentation.

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 Auto Segmentation vary for organs with high inter-patient anatomical variability compared to those with more consistent anatomy, based on the provided DICE scores?; What are the specific known limitations in contouring accuracy, particularly for challenging organs at risk like the chiasma or superior Pharyngeal Constrictor Muscle, and how are users guided to address these potential inaccuracies?; Given the training dataset consisted of 911 CT exams, how diverse was this dataset in terms of patient pathology, image acquisition protocols, and scanner types beyond the information provided for the testing dataset?; How does the Auto Segmentation system handle CT images with significant motion artifacts or poor contrast, and are there any integrated pre-processing steps to mitigate their impact on segmentation accuracy?; What is the typical time saved in the radiation therapy planning workflow when using Auto Segmentation for initial contours compared to fully manual contouring, as suggested by the qualitative preference study or other evaluations?