Company Spotlight: DeepHealth ProstateAI - Standardizing Prostate MRI Detection

Sam: A retrospective study of a new AI tool for prostate MRIs showed that when radiologists used it as an aid, their detection performance metric jumped from 0.626 to 0.724.

Maya: That is a statistically significant bump for a very challenging imaging task. Today we are unpacking the FDA 510(k) summary for DeepHealth ProstateAI, published by the U.S. Food and Drug Administration on April 29, 2026.

Sam: And before we dive in, we want to remind our listeners that our voices are AI-generated, and this episode is a summary of a publicly available FDA document, not company marketing material.

Maya: Let us start with exactly what this device is and why it exists. DeepHealth ProstateAI, which was submitted by Quantib BV, is classified as Radiological Computer-Assisted Detection And Diagnosis Software. In regulatory shorthand, that is CADe/x.

Sam: So it is designed to both detect potential lesions and help diagnose or characterize them. What is the specific clinical problem it is trying to solve here?

Maya: The summary states it is developed to aid trained physicians in the detection and characterization of prostate cancer lesions using MR imaging data. It is intended for adult males aged 40 years and older who have clinical indicators suggesting possible prostate cancer.

Sam: Clinical indicators like what?

Maya: The document lists a few examples: elevated serum PSA, an abnormal digital rectal examination, a family history of prostate cancer, or known genetic markers. These are the patients who are referred for a prostate MRI examination.

Sam: Okay, so this is for patients already in the diagnostic pathway. How does the software actually work on the scans?

Maya: It analyzes bi-parametric MRIs. Specifically, it looks at T2W, ADC, and DWI MR images. It uses artificial intelligence, specifically a convolutional neural network or CNN, to automatically detect candidate regions of interest.

Sam: Wait, let me pause you there, because there is a very interesting workflow requirement mentioned in the indications for use. It cannot just take the raw MRI images from the scanner and do everything itself, right?

Maya: That is a critical point for any hospital buyer. The summary explicitly states that combined with the MR images, the software requires a binary mask containing the prostate gland segmentation.

Sam: Which means the hospital needs to have another cleared device or software tool that already draws the boundary around the prostate gland before DeepHealth ProstateAI can go in and look for the cancer lesions.

Maya: Exactly. The document says it requires that prostate segmentation from a cleared device as input. Once it has that mask and the MR images, it generates segmentations of identified regions and classifies each one.

Sam: Let us talk about those classifications, because this is where the AI really weighs in. How does it score the lesions?

Maya: It uses a categorical system with three distinct levels: low, moderate, and high suspicion. And the summary provides the actual design targets alongside the observed validation performance for these categories.

Sam: I love when they put the real numbers in the public summary. What did the validation show?

Maya: For the low category, the design target was a likelihood of prostate cancer of ≤20%. In their validation cohort, 10% of the low-suspicion lesions were confirmed as prostate cancer.

Sam: So one in ten of the low risk flags still turned out to be cancer. What about the next tier?

Maya: The moderate category had a target of >20% to <80%. They observed that 46% of moderate-suspicion lesions were confirmed as cancer. And for the high category, the target was ≥80%, and they observed an 86% confirmation rate.

Sam: Those are very clean, clear buckets for a radiologist to look at. Green, yellow, red overlays on their screen, basically.

Maya: Yes, the document literally says it outputs files for display in viewers as color-coded overlays using green, yellow, and red. But the summary is very firm that this is intended to be used as a concurrent reading aid.

Sam: Meaning the doctor looks at the AI results at the exact same time they are doing their own review of the scan. It is not meant to replace them.

Maya: Right. The indications for use section says analysis by this device is not intended as a replacement for interpreting prostate abnormalities, and that patient management decisions should not be made solely based on the results.

Sam: Let us shift to the data that built this. If I am evaluating this for my health system, I want to know how much data this neural network actually saw before it hit the market.

Maya: The summary notes a training dataset of approximately 3,200 prostate MRI cases. It says these were collected from six independent institutions across multiple geographic regions.

Sam: 3,200 cases is a decent size. Do we know how many of those actually had cancer?

Maya: The document does not give a direct breakdown of healthy versus cancerous patients in that total. But it does say the dataset provided 883 lesion-level ground-truth labels based on radiologist assessments and pathology-confirmed findings.

Sam: Got it. So we know at least 883 specific lesions were mapped and labeled for the AI to learn from. But training data is one thing. How did they actually test its performance for this FDA submission?

Maya: They ran a multi-reader, multi-case reader study, or MRMC, alongside an independent standalone assessment. Both used a retrospectively collected dataset of 250 prostate MRI exams.

Sam: I always have to point out when testing is retrospective. It means they used historical scans, not a prospective trial where they follow new patients in real time. It is standard for this type of clearance, but it is an important limitation to keep in mind.

Maya: That is a fair point. Of those 250 exams, the summary notes that three cases were excluded due to unreliable cancer localization on imaging. So that left 247 cases for the primary endpoint of the reader study.

Sam: And what was that primary endpoint? This brings us back to the numbers we mentioned at the start of the show.

Maya: Yes. The reader study demonstrated a statistically significant improvement in case-level detection with AI assistance. The average AUC-LROC went from 0.626 to 0.724, with a p-value of p<0.05.

Sam: Which means when the doctors had the AI turned on, their accuracy in finding the lesions went up measurably. The summary also mentioned a standalone study, right? How did the algorithm perform on its own?

Maya: The standalone study achieved a case-level AUC of 0.819. They also tested its segmentation abilities, meaning how well it draws the borders of the lesion. It met all targets, with a mean DSC of 0.535 and a mean MSD of 1.641 mm.

Sam: For our listeners who are not imaging scientists, DSC and MSD are standard metrics for how well the AI drawn boundaries match the human expert boundaries. It seems like it hit the marks the FDA was looking for. But wait, how did it actually get cleared? What is the predicate device?

Maya: As a Class II device, it went through the 510(k) pathway, which means it had to prove it is substantially equivalent to a legally marketed predicate device. In this case, the predicate is ProstatID, cleared under K212783.

Sam: Are they identical in how they work?

Maya: No, they have some distinct technological differences, which the summary outlines clearly. While both analyze T2W, DWI, and ADC sequences, ProstatID uses random forest models, whereas DeepHealth ProstateAI uses a convolutional neural network based on a 3D Retina U-Net.

Sam: Random forest versus deep learning neural networks. That is a generational shift in how the software is built. Are there differences in the output for the user?

Maya: Yes. We mentioned the new device gives low, moderate, and high categorical scores. ProstatID, the predicate, outputs an exam score suggestive of PI-RADS and provides volumetric measurements in PDF reports.

Sam: And there is the workflow difference we talked about earlier. ProstatID directly analyzes the datasets without needing segmentation input. But DeepHealth requires that separate binary mask of the prostate.

Maya: Correct. But despite these differences in algorithm design, input requirements, and presentation, the FDA concluded they share the same intended clinical purpose and do not raise different questions of safety or effectiveness.

Sam: Speaking of safety, what does the document say about the risks of using this software? AI is never perfect.

Maya: The summary includes a benefit-risk discussion. It states that as a computer-aided detection device, there is a risk that identifying suspicious lesions may also lead to biopsies of lesions that are ultimately determined to be non-cancer.

Sam: The classic false positive problem. It might help you find more true cancer, but you might also end up biopsying tissue that turns out to be benign.

Maya: Exactly. But the document explains that detecting cancer has immense value. It specifically notes that finding clinically significant prostate cancer, which it defines as ISUP Grade Group ≥ 2, is the principal objective because that drives morbidity and mortality.

Sam: Because that is the stage where doctors generally start looking at curative treatment.

Maya: Right. But it is also intended to assist in detecting lower grade disease, specifically ISUP Grade Group ≥ 1. The summary notes that detecting Grade Group 1 disease has clinical value because it supports enrollment in active surveillance protocols and helps establish a baseline for monitoring.

Sam: So it helps catch the aggressive stuff early, and it helps map out the slow-moving stuff so they can keep an eye on it. Where is this software actually meant to be installed? Can a small mobile clinic run it?

Maya: The use environment section is very restrictive. It is intended for professional medical imaging environments, specifically listing hospital radiology departments and outpatient imaging centers. It explicitly says it is not intended for use in emergency departments, intensive care units, general practitioner offices, mobile imaging units, or patient homes.

Sam: That makes total sense. You need serious infrastructure for this. It mentions needing orchestration systems and diagnostic-quality displays, right?

Maya: Yes, and the MRI machines themselves matter. The validation covered 1.5 T and 3 T scanners from Philips, GE, and Siemens. So it is built to integrate into standard, high-level PACS-based clinical workflows.

Sam: Before we wrap up, I want to touch on the quality standards. Software acting as a medical device usually requires strict lifecycle management.

Maya: Definitely. The summary lists multiple FDA-recognized consensus standards followed during development, including ISO 14971 for risk management, IEC 62304 for software life cycles, and IEC 62366 for usability engineering.

Sam: Usability engineering is crucial when you are designing an interface that highlights potential cancer. A poorly designed overlay could easily mislead a physician.

Maya: Agreed. They also followed specific FDA guidance for cybersecurity in medical devices, which is essential since this software integrates into hospital networks passing sensitive DICOM data.

Sam: To wrap this up, DeepHealth ProstateAI is a newly cleared AI tool that acts as a concurrent reader for prostate MRIs. It requires an external tool to segment the prostate gland first, but once it has that, it uses deep learning to highlight and categorize suspicious lesions into low, moderate, and high risk tiers.

Maya: And in its retrospective reader study, it demonstrated a statistically significant improvement in how well doctors could detect those lesions, shifting the AUC-LROC score from 0.626 to 0.724. It is an impressive technological step, cleared based on its substantial equivalence to an earlier tool.

Sam: As always, you can read the exact numbers and the full FDA substantial equivalence letter yourself. We have linked the source document right in the show notes.

Maya: And a final reminder that this podcast is for informational purposes only and is not medical advice. Thanks for listening to AI in Medicine - Smart Summaries.