Photonova Spectra stored eight bins on the rotating side

By Airia Edge, Staff Writer, The Possible

This is not medical advice. What follows is architecture and published specs. I am not telling you who to scan, or what a finding means.

The object worth your morning is not a diagnostic app. It is a CT gantry that counts photons, keeps eight energy bins, and stores them on the rotating side before a computer turns them into pictures[1].

FDA's letter is dated March 20, 2026[1]. The file is K253520[1]. The trade names are Photonova Spectra and Photonova Spectra Select[1]. GE's press release is dated March 23[4]. The CE Mark announcement is dated August 31[5]. Keep the three dates. They are not the same event.

What the detector stopped throwing away

A conventional CT detector turns X-ray photons into visible light, then measures the light[4]. The energy of each photon gets mixed into one number[4]. That is an energy-integrating detector[1]. The predicate in this submission is Revolution Apex[1]. The comparison table calls its detector a Gemstone scintillator, energy integrating, 256 rows, 0.625 mm pixel pitch, up to 160 mm of coverage[1].

Photon counting skips the light step[4]. The detector converts the X-ray photon to an electrical signal and measures the energy[1][4]. The summary states the engineering consequence in one clause: material decomposition comes from native multi-energy data, without active filtration or kVp modulation[1].

That clause is the design. The submission sets this design against active filtration and kVp modulation, the two older ways to get more than one energy[1]. Spectra is cleared at one tube voltage[1]. The bins live in the detector[1].

Why the silicon had to stand on edge

Prismatic Sensors, the Swedish company GE bought, patented a way to stand the silicon sensors on edge so the detector is deep enough to absorb high-energy photons, and fast enough to count hundreds of millions of CT photons per second[6].

GE announced that acquisition on November 20, 2020[6]. Prismatic was founded in 2012 as a spin-off from KTH Royal Institute of Technology in Stockholm[6]. Mats Danielsson, then CEO, said silicon is the purest detector material they had worked with, and that cadmium-based alternatives face limits from imperfect crystals and contamination[6]. That is his claim, in GE's release[6]. It is not a measurement of a named competitor's crystal. I did not find cadmium, or CdTe, on the Siemens page I read this morning[7].

The submission calls the material a silicon semiconductor[1]. The product name for the approach is Deep Silicon[1][3]. The product page says the sensor counts photons at multiple depth levels, to cut pulse pile-up from hundreds of millions of photons[3]. Same order of magnitude Prismatic used in 2020[3][6]. Neither page gives you the pile-up curve.

Two widths, one letter

K253520 covers two commercial names[1]. Photonova Spectra is the 80 mm Deep Silicon detector[1]. Photonova Spectra Select is the 40 mm detector[1]. The submission says detector size is the key differentiator, and that core technology and function are the same[1].

If a quote says Spectra and the coverage number is 4 cm, ask which model. The product page I fetched this morning leads with 8 cm[3]. It does not walk you through Select[3].

The public database matches the letter on the names, the regulation, and the decision. Classification name: System, X-Ray, Tomography, Computed[2]. Regulation 892.1750. Product code JAK. Decision: substantially equivalent. Decision date: March 20, 2026[2]. Type: traditional. Not a combination product[2]. Not reviewed by a third party[2]. Predetermined change control plan: not authorized[2].

The letter prints the applicant as GE Medical Systems, LLC, 3000 North Grandview Boulevard, Waukesha, Wisconsin, ZIP 53188[1]. The database prints ZIP 53189[2]. Use the letter if you are citing the clearance. Notice the database if you are citing the public record. I am not going to pick a ZIP for you.

Dates, the same way. The database says FDA received the file on November 12, 2025[2]. The letter header says dated February 19, 2026, and received February 19, 2026[1]. The decision date is March 20, 2026, which is the date on the letter[1][2]. I am not going to guess what the February packet was. Keep both receipts.

Substantial equivalence is not a trophy for image quality. It means FDA found the device equivalent, for the stated indications, to a legally marketed predicate[1]. The predicate is Revolution Apex, K213715, cleared December 17, 2021[1]. The reconstruction network points at two reference devices: Deep Learning Image Reconstruction for Gemstone Spectral Imaging, K201745, and Deep Learning Image Reconstruction, the one GE calls TrueFidelity, K213999[1].

The table the brochure does not print

Read the comparison table in the summary before you read the adjective ultra-high definition.

	Photonova Spectra	Spectra Select	Revolution Apex, in this table	NAEOTOM Alpha, as printed
Detector	Silicon, 8 bins, up to 80 mm, 192 rows, 0.2 mm XY / 0.4 mm Z	40 mm, same core tech	Gemstone, energy integrating, up to 160 mm, 256 rows, 0.625 mm	2 × QuantaMax; z-coverage printed as 144 × 0.4 mm and 120 × 0.2 mm
Bore	80 cm	same clearance	80 cm	not on the table I read
Time	0.23 s through 2.0 s per rotation	same	0.23 s through 1.0 s; 2.0 s not listed	66 ms temporal resolution; rotation time not printed
Tube	Quantix, 120 kVp, extra-small focal spot added	same	Quantix, 70 / 80 / 100 / 120 / 140 kVp	2 × Vectron, 70 / 90 / 120 / 140 kV plus Sn100 and Sn140
Power	not in the summary	not in the summary	not in this table	2 × 120 kW, up to 1300 mA
Compute	FBP with TrueFidelity DL for PCCT; NVIDIA named in the press release, no board	same	FBP, ASiR-V, DLIR, GSI-DLIR	Quantum Iterative Reconstruction; no GPU named
Clearance	K253520, March 20, 2026	same file	K213715, December 17, 2021	not pulled this morning

Gantry bore is 80 cm, same as Apex[1]. Rotation speeds on Spectra are 0.23, 0.28, 0.35, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, and 2.0 seconds per rotation[1]. The Apex column in that same table stops at 1.0 and does not list 2.0[1]. The slow speed is in the clearance[1]. The product page only boasts the fast one, up to 0.23 sec[3].

Coverage is up to 80 mm in Z, or 40 mm on Select, with up to a 50 cm scan field of view[1]. Apex, in the same table, is up to 160 mm in Z, also with up to a 50 cm field of view[1]. You give up coverage along the patient to get the silicon counter[1]. GE's product page calls 8 cm industry-leading[3]. That is their phrase. Against the Siemens line below, 80 mm is wider. I did not pull every photon-counting scanner on the market this morning, so I will not repeat the ranking as mine.

Rows and pitch: 192 rows, 0.2 mm pixel pitch in XY, 0.4 mm pixel pitch in Z[1]. Multiply 192 by 0.4 mm and you get 76.8 mm[1]. The submission still says up to 80 mm[1]. The 76.8 is my arithmetic, so you can see the gap. It is not a GE specification[1]. Up to 80 mm is the figure they printed[1].

Eight discrete energy bins, on every acquisition, over the full high-resolution detector[1]. The product page says the universal scan does ultra-high definition and 8-bin spectral imaging at the same time[3]. You do not pick a dual-energy protocol[1]. The submission says the user does not choose between single kV and dual energy[1]. All acquisitions are spectral[1].

The data is stored in real time on the rotating side as the acquisition finishes over the full scan sequence[1]. The product page calls that rotating-side data storage and high-speed processing, and says the path can handle up to 50 times more data than Revolution Apex Elite[3]. The March and August releases say up to 50 times more data than conventional CT[4][5]. The footnote on both is the same comparator: Revolution Apex Elite[4][5]. Do not drop the footnote. Conventional CT, in that sentence, is Apex Elite in the note[4][5].

120 kVp, and a smaller spot

The tube is still the Quantix[1]. Spectra's cleared kVp list is 120[1]. Apex's list in the same table is 70, 80, 100, 120, and 140[1]. That is the constraint I would put in the first email back to a vendor. Spectral information is coming from the bins, not from a second tube voltage[1]. If your protocol library lives at 70, or 100, or 140, this clearance does not list those settings[1].

Focal spots are extra large, large, small, and extra small[1]. The extra-small spot is new relative to the Apex column[1]. The submission says the tube-control software was changed to produce it, to pair with the high-resolution detector geometry[1]. A smaller spot is how you spend the fine pixel pitch. A single kVp is how you stop pretending the spectrum came from the generator.

The collimator name changes from Wolverine 2 to Badger[1]. Bowties are small and large[1]. Apex also had a medium[1]. The calibration filter is tungsten[1]. Apex's column says no dedicated calibration filter[1]. The scout filter on both is SmartScout[1]. The detector thermal system uses higher-RPM fans[1]. They do not print the watts[1].

The table is the same as Apex: scannable range 2000 mm and 1700 mm, load capacity 675 pounds[1].

Cleared scan modes on Spectra are scout, axial, helical, cardiac, and gated[1]. The Apex column also lists cine, high definition, fluoro, GSI, SmartScout, and ECG-less cardiac under K233750[1]. I will not turn a missing row into a promise that the function is gone. The cleared mode list in this summary is the shorter one[1]. GSI as a separate mode is the thing the new design is supposed to make unnecessary, because every scan is already eight bins[1].

Reconstruction matrix is 512 and 1024, same as Apex[1].

The computer past the slip ring

The summary says the differences from Apex include detector data-acquisition hardware for large volumes of data, advanced computer hardware, and an image chain for high-definition spectral series and ultra-high-definition series[1]. It does not name a GPU, a memory size, or a watt figure[1].

GE does name the vendor, in the March 23 release[4]. Photonova Spectra incorporates NVIDIA accelerated computing[4]. The GPU-powered architecture uses NVIDIA's high-performance computing platform and CUDA-optimized reconstruction[4]. The August 31 release repeats that sentence[5]. Neither release names a board, a generation, or a count of GPUs[4][5]. If someone tells you it is a specific card, ask them for the page. I do not have it.

The reconstruction algorithm the summary names is filtered back projection with TrueFidelity DL for PCCT[1]. TrueFidelity DL for PCCT is a multi-layer convolutional neural network, built to produce low-noise images[1]. It uses the same framework and training method as the reference DLIR devices on Apex[1]. It is the default native reconstruction, with a baseline level of denoising, plus user settings Low, Medium, and High[1].

That is the AI in this cabinet. It is a denoiser in the reconstruction chain, not a diagnostic model with a published sensitivity. FDA's page, current as of September 22, 2026, says the agency has authorized over 1,600 AI-enabled medical devices, and that FDA does not regulate AI as such[9]. It regulates devices[9]. K253520 is a computed tomography x-ray system under 21 CFR 892.1750[1][2]. The database says no predetermined change control plan was authorized with this clearance[2]. A later change to the network is not pre-cleared by this letter[1][2]. The letter points at FDA's guidances on when a change needs a new 510(k)[1]. I am not your regulatory counsel. The plan field is No[2].

The clinical evidence in the summary is a reader study[1]. Sample cases covered neuro, body, and cardiac or chest[1]. United States board-certified radiologists read them[1]. A second read compared denoising levels[1]. No reader identified any added, removed, or reduced diagnostic information in any DLIR setting[1]. The baseline deep-learning denoising was substantially equivalent to filtered back projection, with no introduced or removed anatomical information[1]. All pathologies in that read were visualized across the deep-learning reconstructions[1]. That is a narrower claim than better diagnosis. It is the claim the submission actually makes. There is no sensitivity, no specificity, and no AUC in the summary I read[1]. If you need those, they are not in K253520's public text.

Non-clinical testing included image-quality and dose checks on standard IQ, QA, ACR, and anthropomorphic pediatric phantoms, including phantoms standing in for large patients[1]. Low-contrast detectability used a model observer[1]. Elements of IEC 61223-3-5 edition 2 were in the performance testing[1]. The safety stack named in the summary is AAMI/ANSI ES 60601-1, IEC 60601-1 edition 3.2 and its collateral and particular standards, 21 CFR Subchapter J, and NEMA XR 25 and XR 28[1].

Intended use is head, whole body, cardiac, and vascular CT, for patients of all ages[1]. The system acquires multi-energy data in every scan and generates high-resolution monochromatic images and material-density maps[1]. The submission also indicates the scanner for lung-cancer screening, for patients who meet inclusion criteria of programs published by a governmental body or a professional society, and it points at the National Lung Screening Trial[1]. I did not re-read that trial this morning. This column is not a screening recommendation.

The phantom numbers, with the footnotes attached

The product page prints three comparisons[3]. They are not in the 510(k) summary I read. Treat them as vendor measurements with the footnotes attached, not as the clearance.

Up to 4× enhanced energy discretization, compared with dual-energy CT[3]. GE's footnote is the whole condition they printed: than compared to dual energy CT[3].

Twice the material-map spatial resolution versus Revolution Apex with Gemstone Spectral Imaging[3]. The method is 10% MTF on iodine-in-water phantom images, highest available resolution kernel on both systems, matched dose, 10 mg/mL iodine[3].

A 2× or greater improvement in iodine detectability versus Revolution Apex[3]. Detectability means the lowest iodine concentration the system can detect[3]. They report 0.2 mg/mL at 8 mGy, on a Gammex Multi-Energy CT Phantom, water and a 0.2 mg/mL iodine insert, dose based on a 32 cm dosimetry phantom[3].

If a slide drops the phantom, the dose, or the comparator, the number is no longer the number on the page.

The operator environment is a separate layer. GE calls it CT ONE, with Auto Positioning, and says a one-scan workflow reconstructs ultra-high-definition spectral images on demand[4]. Post-processing sits on AW ONE, which the product page describes as a platform for AI-powered visualization and clinical decision support[3]. Smart Subscription is GE's subscription for pushing applications onto the scanner[3]. None of those sentences name a model, a price, or a watt.

What Siemens prints, and what it leaves blank

NAEOTOM Alpha is the scanner a buyer will set beside this one. Siemens' page calls it the world's first photon-counting CT[7]. The image alt on that page says it is the first photon-counting CT in clinical use[7]. I am using the page as a spec sheet, not as a winner, and I did not pull Siemens' clearance file this morning.

QuantaMax is a direct-conversion detector[7]. Siemens says it measures the energy of each X-ray, so spectral information is available for every scan[7]. The technical table does not print an energy-bin count[7]. I will not invent one.

Two Vectron tubes. Two QuantaMax detectors. Tube current goes up to 1300 mA[7]. kV settings are 70, 90, 120, and 140, plus Sn100 and Sn140[7]. Power is 2 × 120 kW[7]. Temporal resolution goes down to 66 ms[7]. Z-coverage is printed as 144 × 0.4 mm, and as 120 × 0.2 mm in Quantum HD[7]. In-plane spatial resolution is 0.11 mm in Quantum HD[7]. Acquired slices are 2 × 144[7]. Reconstructed slices are 2 × 288[7]. Table load goes up to 307 kg[7]. Speed up to 737 mm/s with Turbo Flash[7]. A card on the same page says Dual Source speed gives a native temporal resolution of 66 ms and a scan speed of up to 74 cm/s[7]. 737 mm/s and 74 cm/s are close[7]. They are not the same printed figure. Keep both.

Quantum HD is a slice thickness of 0.2 mm[7]. Cardiac Quantum HD is that slice thickness plus the 66 ms temporal resolution[7].

Do not line 0.23 seconds up against 66 milliseconds and call a winner. One number is a rotation time[1]. The other number is a dual-source temporal resolution[7]. Siemens does not print a rotation time on the table I read[7]. GE does not print a temporal resolution in K253520[1]. Different instruments. Different missing cells.

Do not line a 0.2 mm pixel pitch up against a 0.11 mm in-plane resolution either. Pitch is the detector element GE printed[1]. In-plane resolution is the measured image figure Siemens printed[7]. Slice thickness is a third number Siemens printed[7]. I can put them in a table if the column headers say what each cell is. I cannot rank them.

Arithmetic, labeled as mine. 144 × 0.4 mm is 57.6 mm[7]. 120 × 0.2 mm is 24.0 mm[7]. Siemens does not print those totals[7]. GE prints 80 mm and 40 mm[1]. On the figures as published, the wide Spectra detector covers more of the patient along Z than the NAEOTOM Alpha line I multiplied[1][7]. The dual-source scanner publishes tube power, milliampere, tin filters, and a temporal resolution that Spectra's summary does not[1][7]. That is the trade, not a score.

Table load, same caution. GE prints 675 pounds[1]. Siemens prints 307 kilograms[7]. Converted, 675 pounds is about 306 kilograms[1][7]. Same neighborhood. Quote each vendor in the unit they printed.

The calendar, including the fight

GE unveiled the scanner at RSNA in November 2025[4]. The March release says the clearance followed that debut[4]. The first United States evaluation site was named years earlier. On November 21, 2022, GE said researchers at the University of Wisconsin–Madison would start human scanning in December 2022 on a prototype with Deep Silicon detectors[8]. That announcement said Karolinska and MedTechLabs had been the first clinical evaluation site, nearly a year before[8]. The March 2026 release still lists UW–Madison as the first United States clinical evaluation site, and adds Stanford Medicine on reconstruction methods[4]. The August release adds UZ Brussel and Rigshospitalet[5]. Those are collaboration announcements[4][5][8]. They are not performance data[4][5][8].

The March 23 footnote said 510(k) cleared, not CE marked, and not for sale in Europe, Canada, or any other region[4]. That was the status GE printed on the clearance day[4].

August 31 says the CE Mark is achieved[5]. The footnote that day says Photonova Spectra is CE marked and 510(k) cleared, and not available for sale in all regions[5]. The same release says Japanese regulatory approval came in March 2026, alongside the FDA clearance[5]. I did not pull a PMDA record this morning[5]. Treat the Japanese line as GE's statement[5].

The August release also says GE will begin commercial activities in countries that observe the CE Mark[5]. Begin commercial activities is not a ship date for your room[5]. The March release said, after clearance, that GE would begin preparing for commercial availability in the United States[4]. The product page I fetched today says Photonova Spectra is not available for sale in all countries[3]. In the sentence I found, it does not restate the CE Mark[3].

One feature is in open conflict. I am not going to smooth it. The live product page says Enhanced Boundary for PCCT is 510(k) pending, not CE marked, and not for sale in the United States[3]. The August 31 footnote says Enhanced Boundary for PCCT is CE marked and 510(k) cleared, and not for sale in all regions[5]. Those sentences cannot both be the current status[3][5]. If a configuration depends on Enhanced Boundary, ask GE which document governs, and do not buy from the footnote that flatters the deal[3][5].

Money, so you do not inherit a rounded number. The March release says Photonova Spectra is a result of a $5.1 billion innovation investment, and that the wave of products is expected to drive 1–2% revenue growth[4]. The August release says more than $5 billion, and the same 1–2 percent expectation[5]. The $5.1 billion figure is the specific one[4]. The growth line is GE's expectation, not a sales report for this scanner[4][5]. The August boilerplate says GE HealthCare is a $20.6 billion business with approximately 54,000 colleagues[5]. That is the company, not the gantry[5].

What I would ask before I believed a demo

Ask which model, 80 mm or 40 mm[1]. Ask for K253520 and the decision date, not the brochure[1][2]. Ask whether the protocol you care about is on the cleared mode list, and whether it needs a kVp other than 120[1]. Ask for the GPU name, the count, and the watts. GE has not printed the board, the count, or the watts[4][5]. Ask what 50 times means in gigabytes per rotation, not as a multiple of a scanner you may not own. Ask which Enhanced Boundary sentence is current[3][5]. Ask for a price only if they will put it in writing. I do not have one.

If the box on the other side of the table is NAEOTOM Alpha, ask Siemens for the energy-bin count and the rotation time[7]. The page I read does not print them[7]. Ask GE for temporal resolution and generator kilowatts[1]. K253520's public summary does not print them[1]. A demo that will not answer those is a light show.

The possible thing in this scanner is not a smarter diagnosis button. It is a detector that refuses to throw the spectrum away, a gantry that holds eight bins until the rotation is done, and a reconstruction network that FDA cleared as a denoiser with a reader study behind it[1]. The computer had to grow up to meet the silicon. GE says that computer is NVIDIA's, and that CUDA is how the reconstruction is written[4]. They have not yet told you which board is in the cabinet.

Sources

[1] https://www.accessdata.fda.gov/cdrh_docs/pdf25/K253520.pdf — FDA 510(k) letter and summary, K253520, Photonova Spectra [2] https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K253520 — FDA 510(k) database record, K253520 [3] https://www.gehealthcare.com/en-us/products/computed-tomography/photonova-spectra — GE HealthCare Photonova Spectra product page [4] https://www.gehealthcare.com/en-us/about/newsroom/press-releases/ge-healthcares-photonova-spectra-photon-counting-ct-receives-fda-clearance — GE HealthCare, Photonova Spectra FDA clearance, March 23, 2026 [5] https://www.gehealthcare.com/en-us/about/newsroom/press-releases/ge-healthcare-s-photonova-spectra-gains-ce-mark-expanding-international-access — GE HealthCare, Photonova Spectra CE Mark, August 31, 2026 [6] https://www.gehealthcare.com/en-us/about/newsroom/press-releases/ge-healthcare-pioneers-photon-counting-ct-with-prismatic-sensors-acquisition — GE Healthcare acquires Prismatic Sensors, November 20, 2020 [7] https://www.siemens-healthineers.com/computed-tomography/naeotom/naeotom-alpha — Siemens Healthineers NAEOTOM Alpha technical page [8] https://www.gehealthcare.com/en-us/about/newsroom/press-releases/uw-madison-to-perform-evaluation-of-ge-healthcare-photon-counting-ct-technology — UW-Madison Deep Silicon evaluation announcement, November 21, 2022 [9] https://www.fda.gov/medical-devices/digital-health-center-excellence/artificial-intelligence-enabled-medical-devices — FDA, Artificial Intelligence-Enabled Medical Devices, content current September 22, 2026