SubtleSYNTH (1.x)

SubtleSYNTH (1.x) is a machine learning-based software device that synthesizes STIR-equivalent MR images of the spine from existing T1-weighted and T2-weighted sequences. Its methodology relies on a convolutional neural network to generate these SynthSTIR images, which have demonstrated interchangeability with conventionally acquired STIR images in non-clinical and reader-based studies.

SubtleSYNTH (1.x) is a software-as-a-medical-device designed for post-processing Magnetic Resonance (MR) images of the spine. Its core function is to synthesize a STIR-like contrast image, termed SynthSTIR, using a machine learning algorithm, specifically a convolutional neural network (CNN). The software operates by taking previously acquired T1-weighted and T2-weighted spine MR images as input. It processes these images in the background, without direct interaction with the MR scanner or a user interface, to generate the SynthSTIR image. This synthesized image is then transmitted via a compatible Medical Device Data System (MDDS) to a Picture Archiving and Communication System (PACS) or other DICOM node for clinical review by radiologists. SubtleSYNTH is intended to be used in imaging centers, clinics, or hospitals, and can operate within a facility's network or remotely, depending on the MDDS configuration, while the device itself is not networked. The performance of SubtleSYNTH was validated through comprehensive non-clinical testing and two key interchangeability studies. Quantitative bench testing on 80 diverse clinical spine MR studies demonstrated that SynthSTIR images maintained consistent relative tissue contrasts compared to acquired STIR images, with a low Root Mean Square Error (0.39) and high cosine similarity (>0.9). Bland-Altman analysis showed no significant bias in pixel values for key tissues (Bone, Disc, CSF, Spinal Cord, Fat). The two interchangeability studies involved radiologists classifying 104 spine cases, presented as either SynthSTIR or acquired STIR images, across various pathologies. Both studies, one presenting only STIR/SynthSTIR (Study A) and the other including additional standard sequences (Study B), successfully met their primary endpoint. The results, analyzed using Generalized Linear Mixed Models and bootstrap methods, indicated low interchangeability percentages (Study A: ~2.1%; Study B: ~0.6%), well within the predefined 10% margin, confirming that SynthSTIR images are interchangeable with traditionally acquired STIR images for diagnostic interpretation across different pathologies and scanner types. The training dataset for the CNN comprised 424 distinct cases, separate from the testing datasets.

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 SubtleSYNTH's SynthSTIR vary across different MRI field strengths (e.g., 0.3T vs. 1.5T vs. 3T) specifically for subtle pathological findings in the spine?; What is the impact on diagnostic confidence and interpretation time for radiologists when using SynthSTIR images compared to conventionally acquired STIR images, particularly in complex cases with multiple pathologies?; Given that the training data's ethnicity is inferred, what steps could be taken to prospectively ensure and validate performance robustness across diverse ethnic populations, and how sensitive is the algorithm to variations not captured in the current demographic inferences?; How does SubtleSYNTH's CNN architecture specifically address potential artifacts or noise amplification that might occur when synthesizing STIR contrast from T1w and T2w images with varying initial quality?; Beyond interchangeability in image classification, what are the quantitative differences in specific measurements (e.g., lesion size, signal intensity ratios for specific tissues beyond the five tested) between SynthSTIR and acquired STIR images, and are these differences clinically significant?