Subtle Medical

Everything Subtle sells is deep learning that changes the image itself: SubtleHD denoises and sharpens MRI, PET and CT acquired faster or at lower dose, SubtleSYNTH generates a STIR contrast from other spine sequences so it need not be acquired, and SubtleALIGN aligns brain MRI automatically. The SubtleEDGE gateway that moves studies between scanner, cloud and archive has no value without the models, and the Subtle-ELITE package is a bundle of the MRI models. Remove the models and nothing is left to sell.

The radiologist reads every study, but reads the model's output as the image, so oversight sits in the protocol. SubtleHD replaces a noisy fast or low dose acquisition with an enhanced one, and SubtleSYNTH offers the option to skip acquiring a STIR sequence and read a generated one instead; in both cases there may be no unmodified comparator on screen. Denoising can alter or hide small lesions on images that look better, not worse.

SubtleHD (PET) lets the radiologist adjust the denoising level, a real control at the point of read. The company does not publish which acceleration or dose reduction factors are validated for which tracer, sequence or anatomy, or whether the original series is kept. An independent SubtlePET validation found degradation at a 66 percent count reduction, which gives a site one published boundary. Structure described, the protocol threshold missing.

The approach and the test data are described; the models and training data are not. Subtle describes convolutional neural network denoising and sharpening, a newer generation in SubtleHD, and deep learning that synthesizes MR contrasts.

The SubtleHD (MR) 510(k) summary characterizes its test set in detail: 16 anatomies, pediatric cases, 1.5T, 3T and low field scanners from several vendors, acceleration levels up to 80 percent, some sites that supplied no training data, and more than half US data. It does not describe the training data. FDA's clearance of SubtleHD (MR) included a Predetermined Change Control Plan, so model updates to that product follow a reviewed plan, though the plan's contents are not published.

No architecture detail, training data description or model card is published, and nothing says whether the model version is recorded in the output. Approach named, update discipline in place for one product, the rest undisclosed.

A short chain with the host named. Subtle builds its models in house; its cloud runs on AWS under a business associate agreement, and images reach it only after the SubtleEDGE gateway removes identifiers. Two further flows are named: SubtleShare returns de-identified customer images to Subtle by default unless the customer opts out, and AiMIFY is sold by Bracco. The site lists no subprocessor beyond AWS, and neither the terms that govern shared images once Subtle holds them nor the regions used are given.

Peer reviewed and prospective evidence exists for MRI acceleration, and the headline figures reach past it. The company's publications list includes a prospective, multicenter, multireader trial in AJNR (2021) showing 60 percent accelerated volumetric brain MRI with quantitative performance preserved, a multicenter multireader trial in AJNR (2023) finding synthetic STIR spine images diagnostically equivalent to conventional STIR, a multicenter, externally validated study of low count whole body PET (npj Digital Medicine, 2021) and a real world PET dose reduction deployment (EJNMMI, 2021).

Independent groups have tested the limits: one validation of the earlier SubtlePET across 110 patients and three scanner models found performance held at a 50 percent count reduction and degraded at 66 percent. The product pages claim up to 80 percent faster MRI and up to 75 percent faster PET with any FDA approved radiotracer, supported by company examples, while published PET evidence covers a handful of tracers.

The SubtleHD (MR) 510(k) summary reports image quality metrics and a blinded reader study of 205 image pairs. The publications page labels its results for investigational use. Strong evidence for part of the claim.

De-identification and data flows are specific; retention is not, and one default matters. The SubtleEDGE gateway removes identifiers before images leave the facility, and the cloud environment is encrypted and run as if it held PHI. Under SubtleShare, customers share de-identified images with Subtle after processing, study by study, and the DICOM conformance statement says the program is part of the contract unless the customer opts out.

The purpose of the shared images and how long they are kept are not stated. De-identification of images is imperfect, especially for burned-in text and head imaging. Safety engineering, meaning how a lost or altered lesion is handled as a safety event, is not published.

Subtle states a scope position: identifiers are removed before anything reaches its cloud. The DICOM conformance statement says the on-premise SubtleEDGE gateway strips PHI and other identifiers before images are sent to the cloud, that Subtle nonetheless treats that environment as if it held ePHI, and that it holds a business associate agreement with AWS; the company is audited annually for SOC 2 Type II and HIPAA. Fully on-premise deployment keeps processing inside the facility.

That is the scope route to B: a plain statement of why the service does not receive PHI, with a service document behind it. No business associate agreement with customers is published, and the website privacy policy does not cover patient images.

SOC 2 Type II is claimed as the company's own and audited annually; the report is not public. Subtle's Compliance Center lists an annual SOC 2 Type II plus HIPAA audit, and its privacy policy adds annual penetration tests and quarterly access and device audits. It also holds ISO 13485 under MDSAP, EU MDR certification and ISO/IEC 42001 for AI management, none of which is an information security attestation. No trust portal offers the report or states its period. The software alters primary diagnostic images inline, yet nothing describes how the integrity of enhanced output is protected.

Each product is positioned as cleared image processing software, and the lead MRI clearance is identifiable. SubtleHD (version 1.x) was cleared on 12 February 2025 (K243250) for noise reduction and sharpening of MRI images of all body parts, and FDA's clearance included a Predetermined Change Control Plan.

The company reports 11 FDA cleared products: SubtlePET (2018) and SubtleMR, SubtleHD (PET) in May 2026, supporting all FDA approved radiotracers, SubtleHD (CT), its first CT product, and SubtleSYNTH (synthetic spine STIR) and SubtleALIGN (brain alignment), which it describes as cleared. SubtleGAD, for reduced gadolinium dosing, is labeled work in progress, and AiMIFY is sold by Bracco. The company holds ISO 13485 under MDSAP and EU MDR certification. The positioning matches the products: they process images for a radiologist to read and do not claim a finding.

A certified AI management system and a cleared change plan stand behind the models; no results by patient group are published. Subtle holds ISO/IEC 42001:2023 certification for its AI management system (valid to July 2028) within its quality management system, publishes an AI policy, and FDA's clearance of SubtleHD (MR) included a Predetermined Change Control Plan.

Nothing compares performance across patient groups: the SubtleHD (MR) summary includes pediatric cases in its test set without separate results. Body habitus is the specific concern for PET: count statistics worsen as body size increases, so denoising does more work in larger patients, while a 2026 study noted a small subgroup above a BMI of 35 and an implementation study found the largest dose reductions in patients with low BMI.

Mechanisms let a reader check the output; nothing stands behind it. SubtleHD (PET) lets the radiologist adjust the denoising level, SubtleHD (MR) can change only under its cleared change plan, and the company is audited annually. The SubtleHD (MR) 510(k) summary (K243250) reports image quality metrics (L1 loss, SSIM, PSNR) and a blinded reader study of 205 image pairs scored for signal to noise, image quality, small structure visibility and artifacts.

Those measure image quality, not an error rate such as lesions lost or altered, so there is no rate to hold the product to. No warranty, indemnity or remediation commitment is published. For software that alters the primary image, the contract terms for a lost or altered lesion are not described, nor whether the acquired series is kept or how such a case would be detected.

Standards based and specified. Subtle publishes a DICOM conformance statement describing the services it supports as a node in the imaging network, and the software sits between any scanner and any PACS through the SubtleEDGE gateway, with named deployments at Mount Sinai, RadNet, Radiology Partners, Middlesex Health and Precision Imaging Centers. The image processing stops at the archive, with no bidirectional integration into a record system named. Nor does Subtle say whether the output records which model version processed the study, or how enhanced or synthetic series are labeled for the reader.

Options and the data path are documented; residency and tenancy are not. The SubtleEDGE gateway sits on premise, strips identifiers before studies go to the cloud, and returns results to PACS; fully on-premise deployment, online or offline, is also supported. The cloud runs on AWS, with encryption at rest and in transit.

The company names no AWS regions and does not say whether processing can leave the country of origin, what the tenancy model is, or what its backup and recovery posture is. The SubtleShare sharing program needs an online deployment, so an offline installation rules it out.

Nothing published on price, and a value proposition that is unusually quantifiable, which makes the silence more conspicuous rather than less.

No rate card, no pricing basis, no statement of whether the model is per scanner, per site, per study or subscription, no term, no minimum and no volume banding. Sales run through enterprise imaging agreements. The company is private, with growth capital raised in 2026, so no filing obligation supplies what the website does not.

What makes this worth more than a one line note is that the return here is measurable in a way most categories are not. The product sells throughput and dose. Both are countable. Published implementation work reports acquisition time reduced by about a quarter, injected activity reduced by more than a third, and average daily patient throughput up by roughly five cases. A department can therefore calculate what the capability is worth to it with real precision, and still cannot compare that against a price without entering a sales process.

The pricing basis is the substantive question and not a detail, because it determines whether the vendor's incentive aligns with the buyer's. A per study fee scales with the volume the software is being bought to increase. A per scanner or subscription fee does not. Two buyers with identical scan volumes can reach opposite conclusions depending on which structure they are offered, and nothing published indicates which exists.

Bounded by modality and acquisition rather than specialty, with validation behind part of the claim. The products cover MRI of all body parts, PET across tracers, CT, synthetic STIR for the spine and brain alignment, on scanners of any make, in hospitals and imaging centers, including pediatric examples.

Validation is published for brain and spine MRI and for FDG and a few other PET tracers, while the claims cover all anatomies and all FDA approved radiotracers; SubtleSYNTH supports the spine only so far. Coverage named, with validation behind part of it.