MRCAT brain

MRCAT brain is a software application designed to generate CT-like images from MR data for radiotherapy treatment planning of brain tumors. It employs an mDixon MR sequence and a convolutional neural network to automatically segment tissues and assign Hounsfield Unit values, providing radiation attenuation information for accurate dose calculation.

MRCAT brain is a software add-on designed for Philips Ingenia, Ingenia Ambition, and Ingenia Elition MR systems. Its primary purpose is to provide clinicians with information on tissue properties, specifically for estimating radiation attenuation during photon external beam radiotherapy treatment planning for patients with primary and metastatic brain tumors. The system utilizes an mDixon MR imaging protocol, acquiring in-phase and water images which undergo automated post-processing where a machine learning-based algorithm, specifically a pre-trained and locked convolutional neural network (CNN), segments bones (skull, upper C-spine, jaw) and the body outline. The core science involves the CNN, trained on matched CT and MRCAT source images, assigning a continuum of Hounsfield Unit (HU) values to tissues based on their fat and water intensities, thereby generating MRCAT images with CT-like density information. These HU values are calibrated using registered CT images, and the system includes sanity checks for positional accuracy and body outline consistency. The MRCAT images can be exported in DICOM format for use in treatment planning systems. Non-clinical performance data indicates high dose accuracy, with simulated doses closely matching CT-based plans (e.g., within 2%/2mm gamma criterion for 95% of patients in the PTV), and geometric accuracy within ±1 mm for a 200 mm sphere. Clinical data further supports that MRCAT brain images are spatially accurate radiation attenuation estimates, aiding in radiotherapy planning.

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 machine learning-based bone segmentation in MRCAT brain (skull, upper C-spine, jaw) compare to manual segmentation or other automated methods in terms of accuracy and processing time, particularly in atypical anatomies or near metallic implants?; What is the range and distribution of Hounsfield Unit (HU) value discrepancies between MRCAT-generated images and actual CT scans across diverse patient populations, and how does this impact dose calculation for different tissue types beyond bone and soft tissue general classifications?; How does the locked nature of the CNN model account for potential variations introduced by future MR hardware or software updates on the Ingenia platform, and are there protocols for re-validating the MRCAT algorithm's performance post-system upgrades?; What specific sanity checks are implemented to ensure the validity of generated MRCAT images, particularly concerning the correct positioning of the imaging field of view and the matching of the MRCAT body outline to the MR, and what is the failure rate of these checks?; Considering the mDixon sequence parameters are mostly fixed, how sensitive is the MRCAT image generation accuracy to minor permissible variations in image stack location or unavoidable patient-specific factors that might influence MR signal intensities?