Company Spotlight: Subtle Medical SubtleHD-CT - Upgrading CT Image Quality With AI

Sam: What if a hospital could upgrade the image quality of its computed tomography scanners without buying any new hardware, just by dropping a deep learning algorithm into its server architecture?

Maya: That is the exact premise we are looking at today. We are unpacking an FDA 510(k) summary for a product called SubtleHD-CT, version 1.x, published by the U.S. Food and Drug Administration.

Sam: Before we dive in, a quick reminder that our voices are AI-generated. This is a summary of a publicly available regulatory document, and we are not giving medical advice.

Maya: Right. And because this is a Company Spotlight episode, we need to be clear that we are evaluating the FDA's public 510(k) summary submitted by Subtle Medical, not the company's marketing materials or website.

Sam: Exactly. We are sticking strictly to the numbers and claims the FDA actually reviewed. So, Maya, what exactly is SubtleHD-CT?

Maya: It is classified as a Medical Image Management and Processing System, specifically a Class II Software as a Medical Device, or SaMD. The document says it consists of a "deep learning artificial intelligence/machine learning (AI/ML) to enhance medical images acquired by computed tomography (CT) scanners."

Sam: And the indications for use? Who is this actually for?

Maya: The intended use is straightforward. The document states it is an AI/ML software intended for the "processing and enhancement of CT images regardless of the manufacturer or model of CT scanner."

Sam: That detail, that it works regardless of the scanner's make or model, is a big deal for health systems. Hospitals usually have a mix of scanners from different eras and different vendors. Getting them all on the same software ecosystem is notoriously difficult.

Maya: It is. And the workflow here is designed to be invisible to the technologists and radiologists. The software is post-processing, meaning it does not directly interact with the scanner itself.

Sam: Wait, so how does it actually get the images if it doesn't talk to the scanner?

Maya: Once a scan is acquired, the technologist sends the study from the scanner to a compatible medical device data system using the standard DICOM protocol. That system makes the images available to SubtleHD-CT. The software enhances the study using an image enhancement algorithm based on a convolutional neural network.

Sam: And then it just routes it back?

Maya: Exactly. It sends the enhanced images back to the picture archiving and communication system, or PACS, for clinical review. The summary explicitly notes that because the software runs in the background, it has no user interface.

Sam: No user interface at all. The radiologist just opens their normal viewer and sees the enhanced images alongside the originals. The document mentions that the interpretation of the output should be "adjunctive to the original reconstructed image."

Maya: That is a crucial point for clinical practice. It does not replace the original scan; it sits alongside it as an adjunct.

Sam: Since this is a 510(k) clearance, that means they didn't have to prove this was a brand-new invention, right? They just had to prove it was substantially equivalent to something already on the market.

Maya: Correct. The legally marketed predicate device they compared it to is called ClariCT.AI, manufactured by a company called Claripi. Both are intended to enhance CT images across different scanner manufacturers, and both operate on DICOM files using pre-trained deep learning models.

Sam: Okay, so how did they prove to the FDA that it actually works as well as the predicate? What did the performance testing look like?

Maya: They ran two main types of testing: quantitative performance validation using physical phantoms, and a reader study using retrospective clinical data. Let's start with the phantoms, which are essentially calibrated physical objects used to test imaging equipment.

Sam: I love phantom testing because it takes all the human variability out of the equation. Which phantoms did they use?

Maya: They used three industry-standard phantoms: the CT ACR 464 phantom, the Digital CatPhan 600 phantom, and the Mercury 4.0 phantom. They scanned these phantoms across multiple vendor scanners, specifically GE, Siemens, and Canon.

Sam: And what exactly were they looking for in these phantom scans?

Maya: They used simulated targets to estimate how well a signal or lesion could be detected. For example, in the ACR 464 phantom, they looked at simulated circular lesions with sizes of 3mm, 4mm, 5mm, 6mm, and 7mm in varying backgrounds like air, polystyrene, acrylic, and bone.

Sam: That is incredibly specific. What about the CatPhan?

Maya: For the Digital CatPhan 600, they evaluated simulated circular lesions of 3mm, 5mm, 7mm, and 10mm at contrasts of 14 HU, 7 HU, 5 HU, and 3 HU respectively.

Sam: So they are really pushing the algorithm to see if it can maintain or improve the visibility of these tiny, low-contrast lesions. Did it pass?

Maya: Yes. The testing encompassed a wide operational range. To give you the exact numbers, they tested slice thicknesses from 0.28 to 5 mm, fields of view from 144 to 500 mm, and pixel sizes from 0.28 to 0.98 mm. They also used the Mercury 4.0 phantom which has 5 cylindrical modules varying from 16, 21, 26, 31, and 36 cm in diameter to represent different patient sizes.

Sam: And the actual endpoints they had to hit?

Maya: For the primary endpoints, the detectability index, known as d', had to be equal to or greater than the unenhanced inputs. It passed. For image noise magnitude, the enhanced outputs had to be statistically significantly lower, with a p-value of < 0.05 using a Student's t-test. It passed that as well.

Sam: What about ensuring the AI doesn't hallucinate or warp the actual tissue densities? That is always the fear with generative AI in imaging.

Maya: They measured CT Number Accuracy as a secondary endpoint. The requirement was that it must remain within ± 5 Hounsfield units of the unenhanced inputs. It passed. They also looked at image non-uniformity, requiring the difference between peripheral and center regions of interest to be < 5 Hounsfield units. It passed that too.

Sam: Okay, so the physics and the math check out on plastic phantoms. But before we get to the human reader study, I noticed a really significant warning in the document regarding the type of reconstruction the software supports. This seems critical for any hospital buyer.

Maya: Yes, this is a major limitation. The document explicitly states that the software's quantitative performance was validated on CT images reconstructed via standard Filtered Back Projection, or FBP. It says the software is engineered to complement FBP-based imaging.

Sam: And the warning?

Maya: It says "use with Iterative Reconstruction (IR) or Deep Learning (DL)-based reconstructions may result in unknown performance."

Sam: Unknown performance. That is massive. Modern scanners often use Iterative Reconstruction natively to lower radiation dose, and newer high-end machines come with their own deep learning reconstruction. So this software is specifically validated for the older, standard Filtered Back Projection method.

Maya: Exactly. The document also notes that performance has not been specifically characterized on images that have already undergone advanced third-party post-processing or specialized non-standard filtering. They recommend applying the enhancement to the direct output of the scanner's primary reconstructed images.

Sam: Alright, so if a hospital is running standard FBP, this could be a great fit. Let's talk about the human data. How did they test this on real patient scans?

Maya: They conducted a reader study utilizing 270 cases of retrospectively acquired human data. These were obtained from various institutions as part of research studies or clinical exams.

Sam: Retrospective is the key word there. They didn't test this live on patients in the emergency room; they ran old scans through the algorithm to see how it performed.

Maya: That is correct. To ensure the dataset represented the actual patient population, they included a wide split of anatomies: abdomen, c-spine, cardiac, chest, chest-abdomen-pelvis, head, head & neck, l-spine, lower-extremity, MSK, t-spine, and upper-extremity.

Sam: Did they just use one type of scanner for the human data?

Maya: No, they used images from Canon, GE Medical Systems, Hitachi, Philips, and Siemens scanner vendors. The patients ranged from 3 to 90 years old, with an even distribution of sex, and included both contrasted and non-contrasted images.

Sam: That is a solid spread. I see they also tracked the radiation doses for these clinical cases. We know lowering dose is a huge priority in radiology.

Maya: They provided a summary of the CTDIvol dose levels. Across all protocols, the minimum dose was 0.90 mGy, which was recorded in the Chest protocol. The average dose across all cases was 15.38 mGy. The maximum dose was 68.50 mGy, found in the Head protocol.

Sam: And the clinical conditions? Did they just test this on healthy patients?

Maya: No, the dataset included a wide range of pathologies. The reports accompanying the images noted conditions like chronic sinusitis, trauma, stroke, head injury, osteosarcoma, dissection, pain, fracture, mass/tumor, aneurysm, calcium scoring, and coronary CTA. Though the document notes some conditions were not reported with the data.

Sam: Okay, so they have this robust, retrospective dataset of 270 cases. Who actually evaluated them, and how?

Maya: Board-certified radiologists evaluated the images. They were presented side-by-side with the original and the SubtleHD-CT enhanced images.

Sam: Side-by-side makes sense. What were they grading them on?

Maya: They used two main endpoints. First, diagnostic utility, which was graded on a 5-point Likert scale. The acceptance criteria was that the enhanced images had to be the same or better than the original. The results showed the average scores across all readers were an average of 3 or better, which constituted a pass.

Sam: And the second endpoint?

Maya: Artifact presence. This was graded on a 3-point Likert scale. The criteria was that the software must not introduce artifacts that could impact diagnosis in any image. The results showed the enhanced images had an average of 2 or better, meaning it passed that endpoint as well.

Sam: So, to zoom out to the boardroom perspective. If I am a health system executive, this 510(k) summary tells me that this software can take standard FBP scans from almost any major vendor, process them in the background without my radiologists having to click any extra buttons, and deliver an image that board-certified radiologists say is just as good, if not better, with less noise.

Maya: Yes, that is the value proposition supported by the testing in this document. It offers a way to potentially standardise image quality across a mixed fleet of older and newer scanners, provided they are using Filtered Back Projection.

Sam: But the flip side is that limitation. If my hospital has already invested heavily in the latest iterative reconstruction algorithms or native deep learning scanners, this document explicitly states that performance with those is "unknown."

Maya: Exactly. The FDA cleared it based on substantial equivalence to the predicate device, not on clinical superiority over the newest native scanner technologies. It is an adjunctive tool, and the independence between the training and test datasets—meaning no shared subjects or images—gives confidence in the algorithm's generalisability for its specific intended use.

Sam: It is a fascinating approach to upgrading hardware through software. That is all for this Company Spotlight on the FDA 510(k) clearance for SubtleHD-CT.

Maya: To read the exact performance metrics and phantom parameters for yourself, you can find a link to the full FDA summary document in our show notes.

Sam: And as always, remember that this podcast is for informational purposes only and is not medical advice. Thanks for listening.