Company Spotlight: Milvue TechCare Trauma - Boosting Fracture Detection by 21.8%

Sam: Welcome to AI in Medicine, Smart Summaries. I am Sam, and along with Maya, our voices are AI-generated. Today we are looking at a clinical study where AI assistance boosted a reader's ability to catch adult fractures by 21.8%. We are unpacking the FDA 510(k) summary for TechCare Trauma, published by the U.S. Food and Drug Administration on January 17, 2025.

Maya: Exactly, and to be crystal clear right from the start, we are reading the FDA's public 510(k) summary for this product, which is made by a company called Milvue. This is not a marketing brochure, and we are not sponsored by them. We are strictly looking at the clearance documents to see what this software actually does for clinicians on the ward.

Sam: So let us get right into it. TechCare Trauma is classed as a radiological computer assisted detection and diagnostic software. Basically, it analyzes 2D X-ray radiographs using machine learning to aid in the detection, localization, and characterization of fractures and elbow joint effusions.

Maya: The main problem it addresses is missed fractures or subtle effusions, especially in busy settings. The document specifies that the intended users are clinicians with the authority to diagnose fractures. That includes primary care, emergency medicine, urgent care, orthopedics, and of course, radiologists.

Sam: How does it actually sit in the workflow? Does a doctor have to upload the images manually to a separate app?

Maya: No, and that is a key detail. It relies solely on images provided by DICOM sources. Once integrated, it automatically receives and processes the images without manual intervention. The results are sent to specified DICOM destinations, so they can be viewed on any compatible viewer or PACS system. It can also be deployed on-premises or on the cloud.

Sam: But the document has a very clear warning regarding how it should be used. It states that TechCare Trauma results are not intended to be used on a stand-alone basis for clinical decision-making. Primary diagnostic and patient management decisions are made by the clinical user.

Maya: Right, it is a safety net. It is an aid, not an autonomous doctor. The AI only sees the pixels in the image. It does not know the patient's history, their lab results, or their clinical presentation.

Sam: Let us talk about the interface, because how an AI presents its findings is incredibly important. TechCare Trauma uses two fixed operating points. One is called DOUBT, and the other is called YES. Walk me through that.

Maya: This is a really interesting design choice for managing clinician trust. The DOUBT status is what they call a high-sensitivity operating point. It triggers when there is a suspicious area or subtle finding. It displays a white dotted-line bounding box around the region of interest. It basically says, look closer here.

Sam: And the YES status?

Maya: That is the high-specificity operating point. If the AI algorithm's confidence level is above that threshold, it displays a solid green line. It means the software considers it a definite or unequivocal finding. And if it finds nothing suspicious, the status is NO, and no bounding box is displayed.

Sam: The system also generates a Summary Image for the whole study. It takes all the images in a study and gives an overall assessment based on the most severe finding. It can say SUSPECTED if there is at least one YES, or DOUBT if there is at least one DOUBT but no YES. Or NOT SUSPECTED if it found nothing.

Maya: And they also built in a safety feature for mixed studies. If a study contains some images that the software can process, but also some images that are out of its scope, it adds an asterisk to the summary status. So you might see SUSPECTED with an asterisk. That tells the clinician that while a fracture was found, there are other images in the file that the AI could not evaluate.

Sam: That is a smart way to prevent a doctor from assuming the entire study was cleared by the AI. Speaking of what it can process, let us look at the testing data. They did standalone bench testing on a massive dataset. The fracture detection test used 4,109 radiographs of U.S. adult patients and 2,872 radiographs of U.S. pediatric patients.

Maya: And the demographics are highly detailed. The adult mean age was 60.5 plus or minus 17.5 years. The pediatric mean age was 12.2 plus or minus 5.6 years. For the elbow joint effusion, or EJE testing, they used 280 adult radiographs and 483 pediatric radiographs.

Sam: They emphasize that these cases were independent of the training data, obtained from at least 4 U.S. states and at least 14 different imaging hardware manufacturers. How did they establish the ground truth for these images to know if a fracture was actually there?

Maya: They used American Board of Radiology certified radiologists with a minimum of 5 years of experience. They had a specific pediatric panel of three radiologists for the kids, and an adult panel of three musculoskeletal radiologists for the adults. Two independent reads per case, and if they disagreed, the third radiologist broke the tie by majority consensus.

Sam: So what were the actual performance numbers from that bench test? For adult fracture detection, the Area Under the Curve, or ROC-AUC, was 0.962. And interestingly, for pediatric fractures, it was exactly the same at 0.962.

Maya: Which is exactly what you want to see. The document states this demonstrates that TechCare Trauma detects fractures in radiographs with similar performances on the adult population and on the pediatric population. For elbow joint effusions, the ROC-AUC was 0.965 in adults and 0.976 in pediatrics.

Sam: I want to dig into the breakdown by anatomical regions from the adult bench testing because the performance actually varies depending on the bone. For instance, the ROC-AUC for the adult clavicle was incredibly high at 0.998.

Maya: Yes, but look at the hip and pelvis. The ROC-AUC for the adult hip was 0.955, but for the adult pelvis, it dropped down to 0.893.

Sam: That is a noticeable dip compared to the rest of the body. The document does not offer a medical explanation for why that is, it just provides the raw numbers.

Maya: Exactly. We can't guess why, but providing that breakdown is crucial for transparency. Clinicians need to know that the AI performed slightly differently on a forearm, which had an AUC of 0.969, than it did on a pelvis.

Sam: They also broke it down by hardware manufacturer. They tested images from Konica Minolta, Siemens, Samsung, Agfa, Carestream, Philips, RamSoft, and Fujifilm. They also tested displaced versus non-displaced fractures. For adult displaced fractures, the ROC-AUC was 0.970, and for non-displaced it was 0.951.

Maya: But Sam, bench testing is one thing. How the software actually changes a clinician's accuracy is what health system leaders care about most. That brings us to the clinical data.

Sam: Right, the multireader, multicase retrospective study. They used 304 adult fracture cases, 256 pediatric fracture cases, 109 adult EJE cases, and 100 pediatric EJE cases. Readers looked at the images with and without the AI assistance.

Maya: The reading scenarios were randomly assigned, and they used a washout period of at least one month between scenarios. That ensures the readers would not just remember the image from the first time they looked at it.

Sam: Let us look at the adult fracture results. The readers' baseline ROC AUC was 0.865. With TechCare Trauma, it improved to 0.955. That is a delta of 0.090, with a p-value of less than 0.001.

Maya: And the sensitivity jump is the number you mentioned at the top of the episode. Reader sensitivity improved from 0.807 to 0.983. That is a 21.8% increase. The specificity only went up slightly, from 0.815 to 0.827, representing a 1.47% increase.

Sam: That tells a very specific story. The software is primarily helping clinicians catch things they were missing, hence the huge jump in sensitivity, without causing them to overcall and ruin their specificity.

Maya: Exactly. For pediatric fractures, sensitivity jumped 19.9%, moving from 0.804 to 0.964. Specificity remained exactly the same at 0.797. For elbow joint effusions, adult sensitivity went up 12.7%, and pediatric sensitivity went up 18.2%.

Sam: Across the board, the AI helped readers detect more true positives. But we need to remember a major limitation here: this was a retrospective study.

Maya: Yes, that is a vital point for anyone buying or implementing this tool. A retrospective study uses historical images in a controlled reading environment. It does not measure how the tool performs prospectively in a live, chaotic emergency department where a physician is managing multiple patients. The summary proves it improves accuracy in a controlled test, which is the standard for clearance, but it is not real-world prospective data.

Sam: Which leads into how this device was actually cleared. The 510(k) pathway is based on proving substantial equivalence to a predicate device. In this case, the predicate is BoneView 1.1-US by a company called Gleamer.

Maya: Both are Class II radiological computer assisted detection and diagnosis software for fractures. Both use supervised deep learning on 2D X-ray images. Both integrate via DICOM. But there are a few stated differences.

Sam: TechCare Trauma excludes spine analysis, which BoneView includes. However, TechCare Trauma expands the indications in other areas. It includes neonates and infants from birth to less than 2 years old for fractures, and it adds elbow joint effusions for patients aged 2 and older.

Maya: The FDA document concludes that these minor differences do not alter the intended use of the device and do not affect its safety and effectiveness when used as labeled. Because Milvue provided performance testing for those specific new populations, the FDA determined it is substantially equivalent.

Sam: Let us talk about the training data for a second. The summary says the training dataset was 95,266 images. The age range was 0 to 102, with a mean of 53.3 plus or minus 26.4. That is a massive volume of data.

Maya: It is, and that volume is necessary to cover all the anatomical areas listed in their indications. The table covers ankle, foot, knee, leg, femur, wrist, hand, finger, elbow, forearm, arm, shoulder, clavicle, pelvis, hip, and thorax including ribs. You need tens of thousands of varied images to build a model across all those structures.

Sam: There is one detail about the outputs that caught my eye. Besides the annotated images, if configured, TechCare Trauma can send back a DICOM Structured Report that lists the status and area for each finding. That seems like a big deal for workflow.

Maya: It is a major workflow consideration. If the data is only on the image layer, a doctor still has to type it out. A structured report means the discrete data points can theoretically be pulled directly into the dictation software or electronic health record. But again, the radiologist or clinician still has to sign off, as it is not for stand-alone decision-making.

Sam: The document also notes under biocompatibility that because it is standalone software, it has no direct or indirect patient contact, so that testing is not required. And for electrical safety, it is not subject to testing.

Maya: Let us circle back to the DOUBT versus YES operating points, because looking at the adult bench testing numbers, they actually give the sensitivity and specificity at both levels.

Sam: They do. For adult fracture detection, at the high-specificity point, which corresponds to YES, the sensitivity was 0.906 and the specificity was 0.900. But at the high-sensitivity point, meaning DOUBT, the sensitivity jumps to 0.940, while the specificity drops to 0.845.

Maya: That perfectly illustrates the trade-off. To catch more subtle fractures and hit that 0.940 sensitivity, the system has to flag more false positives, dragging specificity down to 0.845. That is exactly why they present those boundary boxes as dotted white lines instead of solid green ones.

Sam: It is a very transparent way to communicate confidence. Instead of a binary output, it gives a gradient. The dotted line effectively says, I am not certain, but you should double-check this area.

Maya: Looking at the pediatric elbow joint effusion numbers for that same trade-off, it is even more pronounced. At the high-specificity point, sensitivity was only 0.669, but specificity was an impressive 0.977. When they lower the threshold to the high-sensitivity point, sensitivity leaps to 0.901, but specificity drops to 0.933.

Sam: It is clear why the FDA cleared this based on the testing provided. The substantial equivalence is documented, the retrospective data shows significant performance gains, and the safety measures like the asterisk for out-of-scope images are well integrated.

Maya: Are there any other limitations we should highlight?

Sam: Just the standard regulatory boundaries. The clearance mentions that CDRH does not evaluate information related to contract liability warranties. And FDA clearance does not mean the device complies with other federal statutes. It is cleared strictly for the indications tested.

Maya: Alright, time to wrap up. We have been looking at the FDA 510(k) summary for TechCare Trauma by Milvue. It is a newly cleared AI software that analyzes 2D X-rays to detect fractures and elbow joint effusions. In retrospective studies, it significantly improved clinician sensitivity, jumping 21.8% for adult fractures.

Sam: It uses a YES and DOUBT visual system to communicate its confidence levels, and it was cleared based on substantial equivalence to the BoneView device. It was tested on a robust dataset, though the clinical performance data is retrospective.

Maya: It is a tool designed to reduce missed diagnoses, but it is strictly an aid, not a replacement for clinical judgment.

Sam: That is all for this episode. As always, a link to the full FDA 510(k) summary is in the show notes. Please remember, this podcast is for informational purposes only and is not medical advice. Thanks for listening to AI in Medicine, Smart Summaries, and we will see you next time.