SubtleHD (1.x)

SubtleHD is a software-as-a-medical-device (SaMD) that enhances MRI images by reducing noise and increasing sharpness. It operates on DICOM image files, applies a convolutional neural network-based filtering algorithm, and outputs enhanced images to any DICOM-compatible destination. It is intended for use by radiologists and technologists in medical facilities to process all body part MRI images acquired as standard-of-care or accelerated exams.

SubtleHD uses convolutional neural network-based filtering with a single neural network trained for combined adaptive noise reduction and sharpness enhancement, followed by nonlocal mean denoising and unsharp masking filters. It offers optional configurable enhancement levels and operates on Linux-compatible systems (PC or Mac). The primary predicate SwiftMR uses similar CNN-based filtering but is PC-only; the secondary predicate SubtleMR uses separate neural networks for noise reduction and sharpening with a single fixed enhancement level. All three process DICOM files and deliver enhanced images to PACS or other destinations.

ISO 13485:2016 (quality management system); MDSAP certification; device testing includes software verification and validation (unit, integration, system), standalone image quality metrics (L1 loss, SSIM, PSNR), performance validation with retrospective clinical data, and a reader study with board-certified radiologists using Wilcoxon signed rank tests.

SubtleHD is substantially equivalent to its predicates because all three devices share the same indications for use (denoise and sharpen MRI images across all anatomies and protocols), employ convolutional neural network-based image enhancement algorithms with image-guided optimization, process DICOM-compliant files with similar workflows, and demonstrate comparable performance in reducing noise and enhancing sharpness. Although SubtleHD offers more granular configurable enhancement levels and broader OS compatibility than SwiftMR, and uses a single combined model versus SubtleMR's separate models, these differences do not affect the fundamental safety or effectiveness—performance validation shows equivalent or better denoising and sharpening endpoints, and reader studies confirm no introduction of artifacts.