Six months ago we launched the YSOTOPE early-access pilot program with three oncology centers, two of them active PRRT programs and one focused on PSMA-targeted radionuclide therapy. The goal was straightforward: put the platform into daily clinical use, not a controlled research environment, and learn what breaks and what works when the software meets real departmental workflows.
Across the six months, we processed 74 patient scans covering multiple cycles of treatment across the two indications. This post shares what the data told us, including the parts that surprised us and the parts that confirmed what we already believed from our development work.
What the Variability Data Looked Like
The clearest finding from the pilot, and the one that gave us the strongest validation of the patient-specific approach, was the spread in kidney absorbed dose per GBq of administered Lu-177 across patients. For our PRRT cohort, kidney dose per GBq ranged from approximately 0.4 Gy to 2.3 Gy between individual patients, a nearly sixfold range for the same administered activity. That spread is consistent with what has been reported in the dosimetry literature, but seeing it across a real clinical cohort in your own platform is a different kind of confirmation.
Within-patient consistency across cycles was generally good. Patients who had high kidney dose per GBq on cycle one tended to have high kidney dose on cycle two, with an average cycle-to-cycle coefficient of variation of about 12% for the patients with at least three dosimetry cycles. That within-patient stability is important: it means that a single well-characterized cycle provides genuinely predictive information about how that patient will accumulate dose across subsequent cycles. The pilot data made us more confident in the clinical utility of cycle one dosimetry as a planning input.
Where the Algorithms Performed Better Than Expected
We had calibrated our automated kidney segmentation against manual delineations performed by nuclear medicine physicists during our internal validation work. The reported accuracy in that controlled setting was reasonable, but controlled validation datasets have a way of being cleaner than real clinical data from centers with different reconstruction protocols and CT slice thicknesses.
The pilot data genuinely reassured us on this point. Across the 74 scans, our automated kidney volume estimates agreed with physicist-reviewed segmentations within 8% on average. The cases that fell outside that range were all identifiable in the platform's flagging system: one patient with a horseshoe kidney that the segmentation model handled poorly, two cases with post-surgical anatomy that required manual correction, and three early scans from one center where CT slice thickness was thicker than our calibration range. All were flagged for review; none were silently wrong.
The flagging behavior was actually the most important thing to validate. A dosimetry platform that is accurate 95% of the time but silently fails the other 5% is a clinical risk. A platform that is accurate 95% of the time and loudly flags the other 5% for human review is a clinical tool. We built the latter intentionally, and the pilot confirmed that the flag logic was catching the right cases.
Where We Had to Iterate
The DICOM ingestion layer was the source of more revision than we anticipated. Two of the three pilot centers used camera systems from vendors we had tested in our development environment; the third used a vendor combination we had partial coverage for. In that center's data, calibration factor metadata was stored in a vendor-private DICOM tag that our parser did not initially read. We caught this in the first week because every scan from that center triggered our missing-calibration flag, which meant we had to build and validate a new parser branch before the center could go live with quantitative dosimetry.
We also found that our report format needed revision. The initial version of the YSOTOPE dose report presented absorbed dose for six structures (both kidneys, liver, spleen, and two tumor regions of interest) in a table, alongside cycle-by-cycle cumulative dose and a projected cumulative dose for the planned remaining cycles. In user feedback sessions, clinicians consistently told us that they wanted the key decision, specifically whether the planned activity for the next cycle kept the patient within the dose constraint, to be more immediately obvious in the report. We revised the report to put a green/yellow/red status for the next-cycle kidney constraint on the front page, with the full data table accessible on the second page. Engagement with the dosimetry data improved noticeably after that change.
The Workflow Integration Friction We Did Not Predict
The technical platform worked well. The workflow friction that slowed adoption in the pilot came from a source we had not fully anticipated: scheduling. Post-treatment SPECT dosimetry requires the 168-hour scan to happen approximately one week after administration. In departments where SPECT scanning time is allocated week by week, holding a scanner slot open seven days in advance for a specific patient is not the default operational pattern. Two of our pilot centers needed to create a new scheduling category to ensure that post-PRRT dosimetry scans did not get displaced by routine bone scans or stress tests that could be booked on shorter notice.
This is not a software problem. It is an operational workflow problem. But it is a problem that a dosimetry program has to solve, and we now actively help new pilot centers think through their scheduling model before they go live, rather than discovering the friction after the first few patients.
Limitations We Are Documenting Honestly
The pilot data confirmed the partial volume limitation in small lesion dosimetry that we describe in our platform documentation. For the PRRT patients with liver metastases smaller than 15 mm, our lesion dose estimates carry high uncertainty and we note this explicitly in each report. We are not suppressing this in the summary; we believe the clinical teams deserve to know where the numbers are reliable and where they are approximate.
We also have not yet validated dosimetry accuracy against an independent reference standard (simultaneous blood-based dosimetry, external calibrated source measurement) in the pilot setting. That validation work is planned for the next phase. The current pilot demonstrates operational feasibility and internal consistency; it does not yet constitute independent accuracy validation.
What Comes Next
The three pilot centers are continuing into 2025. We are adding a fourth center focused on a different patient population and therapeutic agent, which will extend the pharmacokinetic diversity in our model training data. We are also beginning work on the cycle-adaptive activity recommendation feature: rather than just reporting absorbed dose, the platform will propose a next-cycle activity adjustment to keep the patient on track for their individual dose target, taking into account remaining kidney budget and tumor dose accumulation observed to date.
That feature is a step beyond dosimetry into closed-loop dose planning, and it requires careful clinical validation before it can be used to drive treatment decisions rather than support them. We expect to begin structured clinical evaluation of the recommendation logic later this year. Until that validation is complete, the platform's role is to give clinicians better information, not to replace their judgment about what to do with it.