Silicon Health Solutions, LLC (dba VoxelMD) (Mohammad (Mo) Matthew Fakhri)

Comment on FDA-2026-N-7874-0001 from Mohammad (Mo) Matthew Fakhri, MD, practicing board-certified neuroradiologist (UCSF neuroradiology fellowship; Washington University Mallinckrodt Institute research-track residency and NIH T32 fellowship) and founder of Silicon Health Solutions, LLC (dba VoxelMD). Full comment with references attached.

Disclosure: my company develops Voxel Vision, a generative system that drafts ultrasound reports for radiologist review and signature. It is not FDA cleared. The controls below are meant for report drafting in every modality; ultrasound is my worked example.

I build generative reporting tools, and I am concerned about how they fail. A draft report is an unusually persuasive prompt: fluent, complete, confident, and one click from signature. A fabricated finding reads exactly like a real one, and an omission makes no sound at all.

The evidence on automation bias is sobering. In an eye-tracking reader study (Taib et al., Radiology 2026;320(1):e252590), when AI missed a cancer, median reader sensitivity fell from 71% unassisted to 39% with AI (P = .002), and readers searched those cases less. Incorrect AI suggestions also degrade experienced readers (Dratsch et al., Radiology 2023). A large evaluation of a chest radiograph report generator found clinically significant errors in the AI report alone in 22.8% of inpatient and outpatient cases (Tanno et al., Nat Med 2025), and report generators invent comparisons with priors that do not exist (Ramesh et al., PMLR 2022).

My central recommendation: recognize grounding, meaning traceability of every generated finding to the image and region that supports it, as a named risk dimension and risk control. Question 1 already lists "traceability of the output to primary sources." In imaging the primary source is the pixels, and grounding turns the reader’s task from "is this report plausible?" into "does this image show this finding?", which is what radiologists are trained to answer. When AI outlined the region it flagged, readers’ false-negative rate fell from 33.0% to 20.7% (Bernstein et al., Eur Radiol 2023).

Grounding is not automatically safe: physicians trust localized explanations even when the AI is wrong (Prinster et al., Radiology 2024). It should reduce a device’s risk only when:
1. Every positive finding cites an image (for CT and MR, a series and slice) and, where feasible, a region; ungrounded findings are visibly marked.
2. Every measurement states its source (on-image measurement, technologist or sonographer worksheet, or not measured).
3. The device checks its citations before display (image exists, laterality matches, image type supports the claim).
4. Pertinent negatives are stated only for structures actually assessed; the rest is labeled "not assessed", so silence is never read as normal.
5. Every comparison ("stable", "new", "increased") cites the prior exam actually reviewed, or states that none was available.
6. Management-driving scores (TI-RADS, O-RADS, BI-RADS, LI-RADS, Lung-RADS, Bosniak) are computed from grounded features, not generated.
7. Grounding accuracy is measured, not merely displayed.

Evaluation (Q9, Q14, Q15, Q17): add grounding precision, grounding coverage, ungrounded-finding rate, comparison accuracy and omission rate to benchmarks for multimodal imaging devices. Test report-drafting devices with reader studies that seed known errors into drafts (fabricated finding, omission, wrong side, wrong measurement, false comparison) and measure catch rates with and without grounding, against unaided radiologists. Methods should be modality-specific: radiographs (regions on a few images; lines and tubes), CT and MR (series, slice and change over time) and ultrasound (multi-image, operator-dependent, with a sonographer worksheet).

Postmarket (Q19, Q22, Q24): every draft is edited by a radiologist before signature, so draft-to-signed differences are a ready-made monitoring signal, supplemented by sample-based expert review. Learning from corrections belongs under a PCCP with scoped, probationary, removable rules. When the foundation model changes, the grounding benchmark and seeded-error set should be re-run before release.

I would welcome the chance to help build an open seeded-error test set and grounding metrics for radiology report drafting, starting with ultrasound and extending to radiographs and CT, with academic and community partners.

Comment on Docket FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper
and Request for Feedback (FDA-2026-N-7874-0001)
Grounding generative output in the source images: a safety control for AI in radiology

Submitted by: Mohammad (Mo) Matthew Fakhri, MD, practicing neuroradiologist; founder, Silicon
Health Solutions, LLC, doing business as VoxelMD (www.voxelmd.com)
October 2026

Disclosure. I am a practicing neuroradiologist and the founder of Silicon Health Solutions, LLC
(VoxelMD), which develops Voxel Vision, a generative system that drafts ultrasound reports for a
radiologist to review, edit and sign. It is not FDA cleared. I describe its design below only as a worked
example that the controls I recommend can be built today. The controls themselves are meant for
report drafting in every modality, including radiographs, CT and MR; I use ultrasound as the example
because it is what I am building first. The concerns I raise come from reading and signing reports
every day.

About the author. I am board certified in diagnostic radiology with a Certificate of Added
Qualification in neuroradiology (American Board of Radiology). I completed neuroradiology
fellowship at the University of California, San Francisco, where I served as a clinical instructor, and a
research-track diagnostic radiology residency at the Mallinckrodt Institute of Radiology, Washington
University in St. Louis, including a year as an NIH T32 postdoctoral fellow in Biomedical Imaging
for Clinician Scientists. My research, before VoxelMD, was in imaging analysis methods and machine
learning for presurgical brain mapping, supported by an RSNA Research and Education Foundation
Research Resident Grant and an ACR-AUR Research Scholar award, following two years of
postdoctoral research at Brigham and Women's Hospital, Harvard Medical School.

Summary
Generative AI that drafts radiology reports is reaching clinical practice ahead of the evidence and the
regulatory tools needed to keep it safe, and I am concerned. A draft report is the most persuasive
prompt ever placed in front of a radiologist: fluent, complete, confident, and one click from signature.
A fabricated finding reads exactly like a real one, and an omission makes no sound at all. New eyetracking evidence shows that when AI is silent about a cancer, radiologists search less and miss it.

I ask FDA to:
1. Recognize traceability to the source images ("grounding") as a named risk dimension and
risk control in the two-axis framework (Question 1), credited only when every finding is
checkable against the pixels and the traceability itself has been validated.
2. Add grounding precision, grounding coverage and the rate of ungrounded findings to the
benchmarking competencies for multimodal imaging devices (Questions 9 and 17).
 3. Require human-AI team testing with seeded errors (fabricated finding, omission, wrong side,
wrong measurement, false comparison with a prior) for report-drafting devices, with the unaided
radiologist as comparator (Questions 14 and 15).
4. Treat silence as an output. A device should state what it assessed and what it could not assess,
so that "not mentioned" is never read as "normal".
5. Require comparisons to cite the prior. A statement such as "stable" or "new since prior" should
cite the prior examination the device actually reviewed, or say that none was available.
6. Keep management-driving scores deterministic. Categories such as TI-RADS, O-RADS, BIRADS, LI-RADS, Lung-RADS and Bosniak should be computed from grounded features, not
written as free text by the language model.
7. Use radiologists' edits as the core postmarket signal, and require the grounding benchmark
and seeded-error set to be re-run whenever the underlying foundation model changes (Questions
19, 22 and 24).

1. Why generative reporting in radiology needs specific safety controls
FDA's Digital Health Advisory Committee executive summary warns that hallucinations "that may
appear to be authentic outputs to users, may present a significant challenge in certain health care
applications where highly accurate, truthful information is critical" [13]. Radiology reporting is one of
those applications. The evidence on how radiologists respond to AI output is sobering.
Automation bias from AI silence. In a reader study published in Radiology in July 2026, Taib et al.
tracked the eyes of ten breast screening readers as they interpreted 60 mammograms with and without
AI prompts [1]. On cancers the AI missed, median reader sensitivity fell from 71% unassisted to 39%
with AI (P = .002), and readers' fixation rate on those cases fell (0.44 vs 0.47 fixations per second, P =
.03). Readers trusted the AI's silence and searched less. Median reading time rose from 25 seconds
with no prompts to 34 seconds with four or more (P = .001), so prompt clutter carries its own cost.
The authors concluded that "the greatest negative impact was observed with FN AI suggestions." Two
editorials accompanied the paper [2, 3].

Automation bias from wrong suggestions. When a purported AI proposed an incorrect BI-RADS
category, the share of mammograms rated correctly fell from about 80% to between 20% and 46%,
and very experienced readers were affected as well [4]. Among 140 radiologists reading chest
radiographs, inaccurate AI predictions harmed performance, and experience and subspecialty did not
reliably predict who was harmed [5].

Errors in generated reports. In a large expert evaluation of a chest radiograph report generator,
22.8% of inpatient and outpatient cases had a clinically significant error in the AI report alone [6].
Report generators have also been shown to invent comparisons with prior studies that do not exist [7],
a failure that matters most in CT and MR follow-up, where much of the report describes change over
time.
These studies concern detection prompts and category labels, not generated reports, and their numbers
should not be transferred directly. But a generated report is likely to exert a stronger form of the same
influence: it supplies the full narrative, the measurements, the impression and an implied
management, in the radiologist's own reporting voice. The discussion paper's two-axis framework
captures the consequence of relying on an incorrect output. For report drafting I urge FDA to give
explicit weight to one more factor: how easily the user can check each statement against the
primary evidence before relying on it.

2. Grounding as a safety control (Questions 1 and 2)
By grounding I mean that every positive finding in a draft carries a link to the image (for CT and MR,
the series and slice) and, where feasible, the region that supports it, and that the reader can open that
evidence in one action. Question 1 already names "traceability of the output to primary sources" as a
candidate dimension [15]. In imaging the primary source is the pixels, and the radiologist is the one
person trained to check them in seconds.

Grounding changes the reader's task from "is this report plausible?" to "does this image show this
finding?". The first question invites automation bias; the second is what radiologists are trained to
answer. Evidence from detection AI supports this. When sham AI outlined the region it was flagging,
readers' false-negative rate was lower than when it did not (20.7% vs 33.0%, p = .04) [8], and adding
saliency maps roughly halved automation bias in a mammography simulation (36.1% to 17.8%) [9].

Grounding is not automatically safe. Physicians trust localized explanations more whether or not the
AI is correct [10], and FDA's draft guidance on AI-enabled device software functions warns that
explainability visualizations "if not well designed and validated for the target user group" could
"significantly mislead users" [14]. A wrong citation can lend false authority to a fabricated finding.
Grounding should therefore lower a device's risk in the framework only when all of the following
hold:

1. Every positive finding is linked to a specific image and, where feasible, a region. A finding that
cannot be grounded is visibly marked as ungrounded and never presented with the same weight
as a grounded one.
2. Every measurement states its source (an on-image measurement, a technologist or sonographer
worksheet, or not measured), so the reader can tell a measured value from a generated one.
3. The device checks its own citations before display: the cited image exists, the laterality matches,
and the image type supports the claim (for example, an ultrasound grayscale finding should not
cite only a Doppler frame, and a CT finding should cite a series and slice that include the
structure).
4. Pertinent negatives are stated only for structures the device actually assessed, and anything it
could not assess is labeled "not assessed". This targets the omission failure that Taib et al.
measured.
5. Every comparison with a prior examination ("stable", "new", "increased") cites the prior study
and image the device actually reviewed; if no prior was available, the draft says so.
 6. Scores that drive management are computed deterministically from grounded features rather than
generated.
7. Grounding accuracy is measured and reported (Section 3), not merely displayed.

In the system I am building, each drafted ultrasound finding cites the numbered image it came from.
The server re-points a measurement's citation to the frame that carries the calipers and corrects leftright mismatches between a finding and its cited images. A finding without a valid citation is shown
to the radiologist as "no image cited". Disagreements with the sonographer's worksheet are flagged
rather than silently resolved, and TI-RADS and O-RADS categories are computed by code, not by the
model. I offer this not as proof of safety but to show that these controls are practical to build now, and
that they can be specified and tested. The same design is meant to carry over to radiographs and CT,
where the citation becomes a region on a radiograph, or a series, slice and region on a CT.

3. Evaluating grounding and the human-AI team (Questions 9, 11, 14, 15 and 17)
For multimodal imaging devices (Question 17), I suggest the benchmark include:
Grounding precision: the share of cited images or regions that, on expert review, actually show
the stated finding.
Grounding coverage: the share of positive findings that carry a valid citation.
Ungrounded-finding rate and fabricated-finding rate, reported separately, using an
established error taxonomy such as GREEN or RadFact [11, 12].
Omission rate, broken down by whether the device declared the structure assessed or not
assessed.
Comparison accuracy: the share of comparison statements whose cited prior exists and supports
the claim.
Measurement fidelity: agreement of reported sizes with on-image measurements or the
worksheet, including the unit and transcription errors already listed in Appendix A.

For human-AI team testing (Question 14), report-drafting devices should be evaluated in reader
studies that seed known errors into drafts (a fabricated finding, an omitted finding, the wrong side,
a wrong measurement, a wrong category, a false comparison with a prior) and measure how often
radiologists catch each one, with and without grounding, against unaided radiologists reading the
same cases (Question 15). Holding the draft content constant across assisted conditions (draft alone,
draft with source links, draft with source links plus explicit flags for missing or conflicting evidence)
isolates the effect of the safety control itself. Errors deliberately seeded for testing should be reported
separately from naturally occurring errors, and should not be used to estimate how often errors occur
in practice. Reading time and, where feasible, eye-tracking measures of visual search should be
reported, following Taib et al. The discussion paper notes that clinical confirmation need not always
be prospective (Question 11). A seeded-error reader study is a proportionate, repeatable way to test
the failure that matters most: an error that survives into the signed report.
Evaluation methods should be modality-specific. Chest radiographs are where report generation
research has concentrated; grounding there means a region on one or a few images, with particular
weight on lines, tubes and device positions. CT and MR are volumetric: a citation must name a series,
slice and region, measurements should be derived from the image geometry rather than generated, and
much of the report concerns change from prior studies, so comparison accuracy should be tested
directly. Ultrasound is multi-image and operator-dependent, and it comes with a sonographer
worksheet that can help catch omissions and measurement errors. The worksheet is itself fallible, so
agreement between a device and the worksheet is not independent confirmation; in every modality the
reference standard should come from independent specialists reviewing the full examination.
Evaluation methods for ultrasound report drafting are the least developed, which is why I start there.

4. Postmarket monitoring and change control (Questions 19, 21, 22 and 24)
Report drafting has a monitoring signal that most devices lack: every draft is reviewed and edited by a
radiologist before signature. The difference between the draft and the signed report is a labeled record
of each error the radiologist caught. FDA should encourage manufacturers to capture it in deidentified form and to track edit rates, grounding precision and seeded-error catch rates over time,
with predefined triggers for review. Edits do not reveal the errors the radiologist missed, and a wrong
draft accepted without changes is exactly the failure that matters, so periodic sample-based expert
review (Question 19) remains necessary and must include unedited reports.
Devices that adapt to a practice's corrections should do so under a Predetermined Change Control
Plan (PCCP) that limits the scope of each learned rule, holds new rules on probation until a second
case confirms them, retires rules that are contradicted, and keeps every active rule visible to and
removable by the practice. When the third-party foundation model changes (Question 24), the
manufacturer should re-run the full grounding benchmark and the seeded-error set before release, and
the PCCP should define how far each result may move before the change needs FDA review.
Foundation model versions now change every few months, so this re-benchmarking has to be routine
and inexpensive rather than exceptional.

5. An offer to help
I would welcome the chance to contribute to this work: a seeded-error test set and grounding metrics
for radiology report drafting, starting with ultrasound and extending to radiographs and CT, developed
with academic and community practice partners and released openly so that manufacturers and FDA
reviewers can use them alike. I can be reached at mo.fakhri@voxelmd.com.
Respectfully submitted,

Mohammad (Mo) Matthew Fakhri, MD
Neuroradiologist; Founder, Silicon Health Solutions, LLC (dba VoxelMD)
mo.fakhri@voxelmd.com | www.voxelmd.com
References
1. Taib AG, Partridge GJW, Phillips P, Maxwell-Armstrong C, Chen X, Hofvind SSH, James JJ, Chen Y.
Automation bias in action: eye tracking of humans reading screening mammograms with and without AI
prompts. Radiology. 2026;320(1):e252590. doi:10.1148/radiol.252590
2. Clauser P. AI influences visual search in mammography. Radiology. 2026;320(1):e261855.
doi:10.1148/radiol.261855
3. Abbasi N, Giess CS. Understanding automation bias in imaging AI: an early career perspective. Radiology.
2026;320(1):e261864. doi:10.1148/radiol.261864
4. Dratsch T, et al. Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions
on reader performance. Radiology. 2023;307(4):e222176. doi:10.1148/radiol.222176
5. Yu F, Moehring A, Banerjee O, et al. Heterogeneity and predictors of the effects of AI assistance on
radiologists. Nat Med. 2024;30(3):837-849. doi:10.1038/s41591-024-02850-w
6. Tanno R, Barrett DGT, Sellergren A, et al. Collaboration between clinicians and vision-language models in
radiology report generation. Nat Med. 2025;31(2):599-608. doi:10.1038/s41591-024-03302-1
7. Ramesh V, Chi NA, Rajpurkar P. Improving radiology report generation systems by removing hallucinated
references to non-existent priors. Proc Mach Learn Res. 2022;193:456-473.
8. Bernstein MH, et al. Can incorrect artificial intelligence (AI) results impact radiologists, and if so, what can we
do about it? A multi-reader pilot study of lung cancer detection with chest radiography. Eur Radiol.
2023;33(11):8263-8269. doi:10.1007/s00330-023-09747-1
9. Pesapane F, Latronico A, Abbate F, et al. Evaluating cognitive biases in AI-assisted mammography
interpretation: a simulation reader study of explainable AI across radiologist experience levels. Eur Radiol.
2026;36(10):7639-7650. doi:10.1007/s00330-026-12666-6
10. Prinster D, et al. Care to explain? AI explanation types differentially impact chest radiograph diagnostic
performance and physician trust in AI. Radiology. 2024;313(2):e233261. doi:10.1148/radiol.233261
11. Ostmeier S, Xu J, Chen Z, et al. GREEN: Generative Radiology Report Evaluation and Error Notation.
Findings of EMNLP. 2024. arXiv:2405.03595
12. Bannur S, et al. MAIRA-2: grounded radiology report generation. arXiv:2406.04449. 2024.
13. US Food and Drug Administration. Executive summary: Total product lifecycle considerations for generative
AI-enabled devices. Digital Health Advisory Committee meeting, November 20-21, 2024.
https://www.fda.gov/media/182871/download
14. US Food and Drug Administration. Artificial intelligence-enabled device software functions: lifecycle
management and marketing submission recommendations. Draft guidance, January 7, 2025.
https://www.fda.gov/media/184856/download
15. US Food and Drug Administration. Considerations for the regulation of generative AI-enabled medical devices:
discussion paper and request for feedback. August 2026. Docket FDA-2026-N-7874.
https://www.fda.gov/media/194242/download