DeepHealth Prostate AI Study Models 18.9% Fewer Biopsies

A paper in European Radiology becomes a commercial talking point

On October 8, RadNet's AI subsidiary DeepHealth announced new multicenter findings suggesting that AI risk scoring could spare roughly one in five men an invasive prostate biopsy — a result that landed with the familiarity of a franchise releasing its sequel. The announcement draws on a retrospective study published in European Radiology in late September, led by researchers across Germany, the Netherlands and the United States, and frames a specific, quantifiable promise: that combining radiologists' MRI assessments with DeepHealth's Prostate AI risk classification could have avoided 149 biopsies, or 18.9% of the biopsied population, while still catching 98.4% of clinically significant prostate cancers111213.

The stakes are real. Prostate cancer is the second most commonly diagnosed cancer in men worldwide, with an estimated 1.55 million new cases in 2024 alone1112. MRI has become the standard triage step after an elevated PSA test, but interpretation is far from settled — some scans sit in the ambiguous PI-RADS 3 zone where radiologists genuinely disagree about whether to proceed to biopsy1318. And biopsies, while diagnostically essential, are invasive procedures that can cause infections or bleeding, occasionally requiring hospitalization, and they sometimes find nothing clinically meaningful at all111220. Any tool that reduces unnecessary biopsies without missing dangerous cancers addresses a genuine clinical bottleneck.

What the study actually found — and what it didn't

The cohort was substantial but selective: 787 men already undergoing clinical evaluation for suspected prostate cancer, scanned between 2014 and 2025, of whom 380 — nearly half — had clinically significant disease, defined as grade group 2 or higher1213. That enrichment matters enormously. A population where almost every other man has confirmed significant cancer is very different from a screening population, and the biopsy savings modeled here cannot be transposed directly onto routine practice13.

The core figures describe a combined decision rule, not an autonomous AI verdict. The 18.9% biopsy reduction and 98.4% sensitivity describe what would have happened if AI risk scores had been added to radiologists' assessments of equivocal scans — a simulated counterfactual, not an observed change in care13. No biopsies were actually withheld during the study, which means the research cannot, by itself, establish that deferring biopsy on the AI's advice is safe in everyday practice13. Senior author Francesco Giganti, a professor of radiology at University College London, was careful to frame it the same way: AI as "another piece of evidence," with "the final medical decision remaining with the clinical team"1213.

The detailed performance picture is more textured than the press release headline. At the patient level, the AI's sensitivity for clinically significant cancer was 97.6% versus 92.6% for standard PI-RADS scoring, comfortably passing the study's noninferiority test13. But at the lesion level — the granularity that matters when a radiologist is deciding where to place a needle — the AI's sensitivity was 78.8% versus 88.8% for radiologists, and that noninferiority test failed outright13. The reading worth committing to: this AI is a triage and confidence tool that helps decide whether to biopsy, not a detection tool that outperforms a radiologist at where the cancer is. That distinction will define how hospitals actually deploy it.

From journal to clinical workflow

What makes this announcement more than a journal press release is that DeepHealth's Prostate MR Solution is already cleared and shipping. The Prostate AI component received FDA 510(k) clearance and CE marking in May 2026 and is commercially available, with the broader solution carrying CE Mark Class IIb under the EU MDR in Europe13199. The platform integrates automated lesion detection and risk classification, gland segmentation with PSA density calculation, and PI-RADS-compliant structured reporting, and connects to more than eleven fusion biopsy systems to eliminate manual data transfer117.

For hospitals and imaging networks weighing AI deployment, that integration story may matter more than the sensitivity statistics. Real-world deployments cited by the company report a 27% improvement in lesion detection, a 65% reduction in inter-radiologist segmentation variability, and a 37% reduction in workflow time for biopsy-recommended cases1916. An independent case study of two outpatient radiology centers found read-time reductions of 14.1%, with eight of nine radiologists getting faster, and 232 hours saved across the biopsy workflow16. Reducing variability between readers is a quieter benefit than cancer detection, but in busy practices it is arguably the more deployable one — a PI-RADS 3 read that varies by radiologist is exactly the ambiguity the AI is designed to stabilize.

The regulatory and competitive backdrop

This is also a story about a radiology AI market that has moved decisively from science project to line item. RadNet's Digital Health segment — the umbrella for DeepHealth — reported $32.4 million in second-quarter 2026 revenue, with annual recurring revenue reaching $105.5 million at the end of June, nearly double its year-earlier level13. The company has been assembling a portfolio through acquisition, including its $270 million purchase of French AI firm Gleamer in March, and now spans breast, neuro, chest, musculoskeletal, thyroid, abdominal and prostate imaging on a single cloud-native operating system, DeepHealth OS712. Notably, Prostate AI is part of the UK's TRANSFORM program, the largest prostate screening initiative in that country, which aims to analyze over 100,000 scans9.

GE HealthCare is publicly working on competing prostate MRI lesion detection, so DeepHealth's biopsy-triage evidence arrives into an active contest for imaging AI mindshare13. And the precedent landscape is encouraging: a 2024 retrospective study from the German Cancer Research Center using a similar combination of deep learning and radiologist assessment suggested nearly half of biopsies could theoretically have been avoided without missing a meaningful number of tumors — a more aggressive figure than DeepHealth's 18.9% — while its authors, notably, reached the same caveat: prospective trials must confirm the benefit before implementation1720. A 2026 UK multicenter study of an AI decision support system integrating PI-RADS, PSA density and deep-learning scores reached an identical conclusion, calling explicitly for prospective validation within real-world clinical workflow before clinical adoption18.

The honest bottom line

Both the company and the independent coverage converge on one reading: the evidence is promising, it is regulatory-sanctioned, and it is not yet proof. The 510(k) pathway clears a device as substantially equivalent to existing tools — it does not require demonstration that an AI-guided biopsy deferral strategy improves outcomes, which is precisely what this study models but cannot show1318. The gap between a retrospective simulation and a prospective trial in routine care is the entire remaining question for AI in this diagnostic pathway.

For clinicians, the practical takeaway is not that AI will replace biopsy judgment but that it may soon sharpen it at the margins — the equivocal PI-RADS 3 case where a second opinion would change the call. For the healthcare AI sector, DeepHealth's announcement shows the playbook maturing: publish in a serious journal, secure dual FDA/CE clearance, integrate into the fusion-biopsy install base, join a national screening program, and let recurring revenue tell the story. The prostate AI debate is no longer about whether the algorithms work on archive data. It's about whether they hold up when the biopsy needle stays in the tray.