Claire OCT System

Performance Metrics

Clinical Evidence

The pivotal evidence is a prospective, multi-centre, randomised, two-arm trial (ClinicalTrials.gov NCT05113927, IDE G210177) run at 11 US sites between December 2021 and September 2024 in women undergoing breast-conserving surgery for Stage 0–III invasive ductal carcinoma and/or ductal carcinoma in situ; lobular primaries, multicentric and bilateral disease were excluded. Of 661 consented and 615 enrolled, 613 form the safety set (133 roll-in training, 165 Part A on a developmental algorithm, 317 Part B on the locked commercial algorithm). Randomisation was 2:1 device to control and was revealed only after the surgeon declared the standard-of-care (SOC) lumpectomy complete; the control arm served to keep surgeons blinded during SOC and was not used for the primary analysis. The primary effectiveness analysis is a within-arm before-and-after comparison in the 206 Part B device-arm patients of the full analysis set, so the design is prospective and randomised in enrolment but its headline result is not a between-arm comparison; this entry grades it prospective, not RCT, for that reason. Primary endpoint, per the SSED: after SOC assessment alone, 35 of 206 patients (17.0 percent) still had at least one unaddressed positive margin by NCCN and SSO/ASTRO criteria. After SOC plus Claire the count was 28 (13.6 percent): device-aided shaves fully cleared 7 of the 35, an absolute reduction of 3.4 percent (95 percent CI 1.38 to 6.88) and a relative reduction of 20 percent, which exceeded the pre-specified 1 percent performance goal on a one-sided exact binomial test (p=0.0050, alpha 0.025). The SSED's broader clinical-benefit tally adds two patients whose residual disease was detected but not excised (one with no further tissue to take, one protocol violation), six whose device-directed outer shave contained disease behind a pathology-negative interior margin, and four with close margins of 1 mm or less correctly flagged, for 19 of 206 patients in total. The 40 percent figure that circulated in press coverage (14 of 35) is a subject-level secondary measure of residual disease identified with device aid, not the primary endpoint, and the p-value does not attach to it. Margin-level device-aided decisions on 1,335 assessable margins (SSED Table 7) had [[sensitivity]] 24.62 percent (95 percent CI 14.47 to 36.84), [[specificity]] 91.34 percent, positive predictive value 13.22 percent, negative predictive value 95.95 percent and accuracy 88.1 percent: 16 of 65 residual diseased margins were found, 49 were missed, and 105 flagged margins carried no disease. The SSED is explicit that these are measures of clinical decisions after SOC, not standalone performance. A post-hoc comparison on 1,735 margins put SOC alone at 48.7 percent sensitivity and 72.1 percent specificity, and SOC followed by device aid at 61.5 and 65.8 percent. Standalone ImgAssist 2.0, evaluated post hoc on 1,850 pathology-assessed margins with any AI-flagged cluster counted as a positive margin, reached 88 percent sensitivity and 25 percent specificity at the default 0.75 threshold and 45 and 70 percent at the 0.925 "Show Less" threshold (Table 12). Device use added a mean 0.5 false-positive shaves per patient and a mean 2.8 cm³ of tissue (4 percent of excised volume); 19 of the 206 patients (9.2 percent) went on to a re-excision lumpectomy during follow-up. The peer-reviewed record is a single company-authored development study in Life (December 2023, PMID 38137941): a convolutional-neural-network classifier trained on 585 wide-field OCT margin scans from 151 subjects and tested on 155 pathology-confirmed margins (31 positive, from 29 patients), reaching an area under the ROC curve of 0.976 with sensitivity 0.93 and specificity 0.98 and identifying 96.8 percent of pathology-positive margins. Those are development-set figures on a curated held-out set; the pivotal standalone figures above are the ones that describe the approved device on trial margins. No peer-reviewed publication of the pivotal trial was located for this review, and the SSED lists the algorithm's training data as unpublished. AurisMark reasoning: the device meets the one-mark criteria (a regulatory approval and published evidence on file). It does not clear the two-mark bar on honesty-over-generosity grounds: the prospective evidence is public only through the SSED, the one peer-reviewed paper is a company-authored development-set validation, the margin-level device-aided sensitivity is under 25 percent, the standalone specificity at the default threshold is 25 percent, and the SSED records no marketing anywhere before approval, so there is no documented uptake. Re-evaluate on peer-reviewed publication of the pivotal trial or on independent post-market data.

Inside the Auris+ Listing

Regulatory Approvals

Safety Record

The SSED safety set is 613 participants (control 163, device 317, roll-in training 133). Adverse-event incidence was comparable across groups (30.7, 28.4 and 27.8 percent respectively) and none was reported as related to the device; procedure-related events ran at 23 to 25 percent in every group. Seven serious adverse events occurred, none attributed to the device, and no unanticipated adverse device effects were reported. The device never touches the patient: it images excised tissue with low-power near-infrared light. The safety cost that the SSED does quantify is over-excision. Device use added a mean 0.5 [[false-positive]] shaves per participant (SOC alone 2.2, SOC plus device 2.7), and 67 of 206 participants (32.5 percent) had at least one device-directed false-positive shave; the trial capped device-directed shaves at two per orientation and six per patient. Device shaves averaged 0.6 per patient and 2.8 cm³ against 14.7 cm³ for SOC shaves, and BREAST-Q satisfaction scores showed no significant difference between arms. The other side of the ledger is the [[false-negative]]: device-aided decisions missed 49 of 65 residual diseased margins, and an unmarked margin is not a pathology verdict. Final margin status is set by permanent-section histopathology days later, so a miss at the point of care is corrected by pathology at the cost of a potential return to theatre. No recalls, warning letters or MAUDE signals for Perimeter's OCT systems were located; the approved AI configuration has been on the market only since March 2026 and carries no independent post-market performance data yet.

Intended Use & Indications

Claire (developed and trialled as the Perimeter B-Series OCT with ImgAssist AI 2.0) is used in the operating room immediately after the surgeon excises the lumpectomy specimen. The system acquires wide-field optical-coherence- tomography volumes of the specimen surface, scanning an area up to 8.7 cm in diameter on specimens up to 10 cm across, to a depth of at least 2.2 mm with axial resolution of 15 µm or better and lateral resolution of 50 µm or better, per the verification specifications in the FDA summary of safety and effectiveness data (SSED). The ImgAssist deep-learning algorithm classifies regions of interest in each cross-section as suspicious or non-suspicious and highlights the suspicious ones; the surgeon can run the overlay on, on with a higher threshold ("Show Less"), or off, and reviews the flagged regions concurrently with the images themselves and with the standard intraoperative toolkit (palpation, specimen radiography, intraoperative pathology, ultrasound, radiofrequency spectroscopy, cavity shaves) before deciding whether to shave additional tissue. The device is adjunctive by labelling and by regulation. Its contraindications state that it is not to replace histopathology and is not for diagnosis; the labelling frames it as a tool to rule in disease that standard care may have missed, not to rule out positive findings from standard care. In the pivotal trial, standard-of-care assessment had to be completed before the surgeon looked at Claire images; the SSED notes that real-world use may be concurrent, so the trial measured a narrower use than the approved one.