Jazz

Jazz is a Software as a Medical Device employing a semi-automatic, machine learning-assisted pipeline for the labeling, visualization, and volumetric quantification of segmentable brain structures from MR images in patients with multiple sclerosis or brain metastasis, always ensuring physician oversight and correction capabilities. The device's science is supported by performance data demonstrating its accuracy against ground truth and its reproducibility in generating consistent radiological reports.

AI Medical AG's Jazz is a Software as a Medical Device (SaMD) designed to assist physicians in analyzing MR images of the brain. Its primary purpose is the labeling, visualization, and volumetric quantification of segmentable brain structures, specifically for patients with a known diagnosis of multiple sclerosis (via a dedicated MS pipeline) or brain metastasis (via a metastasis pipeline), culminating in the production of a radiological report. The system functions as a semi-automatic tool, processing MR images through steps such as co-registration, contrast recognition, and optional pre-annotation of abnormalities using machine learning algorithms. A key aspect of its methodology is that physicians retain full control, with the ability to review, accept, correct, or entirely bypass the software's pre-annotations before the final electronic report (available in txt, pdf, or DICOM formats) is generated, which can also be further annotated by the physician. The performance of Jazz was rigorously validated through accuracy and reproducibility experiments using a total of 344 subject datasets, which included healthy individuals as well as patients with multiple sclerosis and brain metastasis. Accuracy was evaluated against manually labeled ground truth volumes, with machine learning models trained on fully segregated datasets using gold standard human expert opinion. Specific acceptance criteria, derived from literature, were met for voxel-wise sensitivity (≥40%), voxel-wise specificity (≥95%), lesion-wise Dice score (≥0.5 for MS models), lesion-wise true positive rate (≥60% for high sensitivity model), and lesion-wise false discovery rate (≤50% for high specificity model). Additional accuracy tests for anatomy localization and image co-registration also passed predefined targets. Reproducibility was demonstrated through process-reprocess experiments showing identical lesion counts, total lesion volumes, and report generation, confirming Jazz's reliability.

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 Jazz's machine learning algorithms for lesion pre-annotation compare across different MRI scanner manufacturers, field strengths, and acquisition protocols?; What is the observed inter-rater variability among physicians when correcting the segmentations provided by Jazz, and how does this variability influence the final volumetric quantification and clinical interpretation?; Can the Jazz system be adapted or extended to segment and quantify brain structures or abnormalities relevant to other neurological conditions beyond multiple sclerosis and brain metastasis, and what would be required for such an adaptation?; What are the specific computational requirements (e.g., processing time per case, necessary hardware specifications) for running the Jazz software, particularly its pre-annotation algorithms, in a typical clinical setting?; How does the 'semi-automatic' workflow of Jazz, which necessitates physician review and potential correction, impact overall diagnostic workflow efficiency and turnaround time compared to fully automated segmentation tools or traditional manual methods?