LumiNE US; Lumi

LumiNE US is a software device designed for pre-operative surgical planning by converting 2D MRI and CT medical images from adult patients into interactive 3D models, viewable with augmented reality, to enhance anatomical and pathological insights for pre-diagnosed conditions. It employs machine learning algorithms (nnUnet) for semi-automatic segmentation of contrast-enhanced T1 MRI scans (for tumor, skin, brain, ventricles) and threshold-based segmentation for CT scans, with its MRI tumor segmentation performance validated for specific tumor characteristics using metrics like Dice Similarity Coefficient and Hausdorff Distance against expert-defined ground truth.

LumiNE US is a software system developed for enhancing pre-operative surgical planning by visualizing medical images. Its primary function is to convert 2D patient imaging data, specifically from Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) scans of adult patients, into detailed 3D models. These models, along with original 2D images, can be viewed in augmented reality using a wireless Head-Mounted Display (HMD) or on a PC monitor, providing (neuro)surgeons, (neuro)surgical residents, or qualified medical professionals with improved insights into patient anatomy and pathology. The system is intended for informational purposes in surgical preparation, specifically for previously known and pre-diagnosed conditions such as Glioblastoma, Meningioma, or Metastasis, and is not designed for diagnostic use. The software utilizes distinct methods for image processing: contrast-enhanced T1 MR scans undergo semi-automatic segmentation for known tumors, skin, brain, and ventricles, driven by machine learning algorithms (specifically nnUnet, forming the T1cSF function), while CT scans use threshold-based segmentation. The MRI T1 tumor segmentation is applicable only for single, pre-diagnosed intracranial contrast-enhancing tumors within specific size parameters (2.0cc-100cc volume, 15mm-75mm diameter). Extensive performance testing, including validation of the machine learning algorithms against an independent U.S. dataset truthed by neurosurgeons, demonstrated clinically acceptable accuracy using Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (95% HD) metrics. This validation confirmed the T1cSF's non-inferiority to existing datasets and its efficacy within defined constraints for segmenting brain, skin, tumor, and ventricles, alongside positive results for manual segmentation, HMD visual quality, DICOM handling, and overall usability.

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 the semi-automatic segmentation algorithms (T1cSF) vary across different MRI scanner manufacturers and imaging protocols not explicitly detailed in the validation dataset?; What is the learning curve for surgeons to effectively utilize the Augmented Reality features and 3D model manipulation tools for pre-operative planning?; Beyond the reported Dice Similarity Coefficient and 95% Hausdorff Distance metrics, what is the impact of segmentation inaccuracies within the acceptable limits on actual surgical planning and decision-making?; How does the system handle anatomical variations or atypical pathologies that fall outside the specific training data parameters for the machine learning algorithms, particularly for structures beyond tumors, skin, brain, and ventricles?; What are the specific computational requirements and processing times for converting 2D images to 3D models, particularly for complex cases, and how might this impact workflow integration in a time-sensitive clinical setting?