MRI Planner

MRI Planner is a stand-alone software medical device that utilizes pre-trained machine learning models to generate synthetic CT images from MRI data for radiation attenuation estimation, and to automatically derive contours of specific healthy anatomical structures for radiotherapy treatment planning. Its purpose is to enhance the efficiency and accuracy of the planning process for photon external beam radiotherapy in adult patients with cancers in the brain, head-neck, and pelvic regions, with performance validated through dosimetric and geometric accuracy testing.

MRI Planner, developed by Spectronic Medical AB, is a software-only medical device intended to aid in radiotherapy treatment planning for adult patients with specific cancers in the brain, head-neck, and pelvic regions. The software processes images from MRI systems to achieve two primary goals: first, to generate synthetic CT (sCT) images that provide information on tissue properties for radiation attenuation estimation in photon external beam radiotherapy planning, specifically for pelvis, brain, and head-neck regions; and second, to automatically derive contours of healthy anatomical structures (bladder, colon, and femoral heads for prostate cancer patients only) to assist in their localization and definition. It is explicitly stated that MRI Planner is not designed to automatically contour tumors or tumor clinical target volumes. The core technology relies on pre-trained machine learning models, with the sCT generation model trained on a dataset of 244 patient MR and CT images, and the auto-contouring model trained on 175 patient MR images with expert manual contours. The system operates on a standard x86-64 compatible system with a CUDA capable NVIDIA GPU and requires Ubuntu Linux 18.04. Performance evaluations demonstrated the MRI Planner's dosimetric accuracy and contouring capabilities. Bench tests comparing sCT with conventional CT showed minimal mean target dose differences (e.g., 0.02% ± 0.31% for pelvis) and ensured that 99% of cases had non-target sCT-CT dose differences below 1.0 Gy or 5% of the CT-dose, with no cases exceeding these limits. Gamma evaluations indicated high pass rates, with 100% of cases passing high-dose gamma criteria for all anatomical regions (average pass rates >99.8%). The auto-contouring function for prostate cancer patients (bladder, colon, femoral heads) was assessed against manual delineations using Dice Score (DSC) and 95% Hausdorff Distance (HD), achieving average DSCs of 0.95 (bladder), 0.90 (colon), and 0.96 (femoral heads), and average 95% HDs of 2.69mm, 4.96mm, and 2.04mm, respectively. These tests were conducted using datasets from multiple hospitals, diverse patient demographics, and various MRI scanner models.

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. What specific MRI sequences were utilized for training the machine learning models for synthetic CT generation and auto-contouring, and how does the device's performance vary when processing images from different input sequences?; How does the MRI Planner's performance in generating synthetic CT images and anatomical contours compare when applied to MRI data from scanner manufacturers, field strengths, or patient populations not extensively represented in the original training and testing datasets?; What are the identified potential failure modes of the machine learning algorithms, particularly in cases presenting with atypical anatomy, significant image artifacts, or post-surgical changes, and what mechanisms are in place to alert users to suboptimal or potentially erroneous outputs?; For the auto-contouring feature, what was the inter-observer variability of the 'expert manual contours' used for training the model, and how does this variability compare to the differences observed between the software's automated output and a single radiation oncologist's delineation in practice?; What is the typical processing time required for generating both synthetic CT images and automated contours for a standard patient dataset (e.g., pelvis, head-neck), and how does this impact integration into existing clinical radiotherapy planning workflows?