The question we get asked most often by nuclear medicine physicists who want to implement patient-specific dosimetry is not "does it work." They already know the physics works. The question is: "what do we actually need to build, and how long will it take?"
The honest answer is: more than most people expect, and the difficulty is rarely in the dose calculation itself.
The Pipeline Problem, Not the Physics Problem
Dosimetry is a multi-step pipeline. Measured absorbed dose requires: a calibrated SPECT acquisition, a reliable quantitative reconstruction, organ and tumor volume segmentation, extraction of time-activity curves from serial images, fitting of a clearance model to those curves, numerical integration to get cumulated activity, and application of dose-kernel convolution or S-value lookup to convert cumulated activity to absorbed dose in gray.
Each step has a failure mode. A pipeline with five steps where each step has a 95% chance of producing clean output still fails roughly 23% of the time on any given patient, before you account for patient-level issues like motion artifacts, off-cycle imaging, or inadequate tracer uptake. The fragile part is not the final calculation. It is the data plumbing between steps.
DICOM Ingestion and Quantitative Prerequisites
SPECT data arrives from the gamma camera as a DICOM dataset. But "SPECT DICOM" is not a single thing. Different vendors store reconstruction parameters, attenuation maps, and calibration factors in vendor-specific private tags or in separate files that may or may not accompany the primary dataset. A platform that reads only standard DICOM tags will silently produce wrong numbers when fed data from certain camera models.
When we built the YSOTOPE ingestion layer, we had to handle at least four different vendor SPECT reconstruction conventions before we could trust that the pixel values we were reading were genuinely in Bq/mL and not in arbitrary counts or some intermediate unit. This is not glamorous work, but it determines whether the downstream physics is real.
Quantitative SPECT also requires that the acquiring institution has performed camera sensitivity calibration. That sensitivity factor, typically expressed in counts per second per megabecquerel, is the conversion coefficient between the acquired image values and absolute activity. If the institution has not established this number against a traceable reference standard, no amount of clever modeling downstream can rescue the absorbed dose numbers.
Organ and Lesion Segmentation
For dosimetry to be clinically actionable, you need absorbed dose values per structure: kidneys, liver, spleen, tumor lesions. That requires segmentation of each structure in each SPECT/CT dataset.
Manual segmentation is the accuracy floor and the throughput ceiling. A nuclear medicine physician who segments both kidneys on a post-treatment SPECT/CT study might spend 20 to 30 minutes on that single task, before any dose calculation begins. Multiply that by three or four serial acquisitions per cycle and four treatment cycles, and manual segmentation alone becomes the clinical bottleneck.
The approach we use at YSOTOPE relies on CT-based automatic segmentation for organs at risk, with threshold-based or gradient-based methods for high-uptake tumor lesions, followed by a physics-informed consistency check: does the sum of organ activity estimates come reasonably close to the injected activity corrected for physical decay? If not, a flag is raised for review.
We are not claiming that automatic segmentation is as accurate as expert manual delineation for every structure on every patient. It is not. What we claim is that semi-automatic segmentation with a physics consistency check is faster than purely manual segmentation and fails loudly when it fails, which is better than failing silently.
PK Fitting and Cumulated Activity
Once you have organ volumes and time-activity data from three or more serial SPECT time points, the next task is fitting a clearance model. For most radiopharmaceuticals used in theranostics, including Lu-177 DOTATATE and Lu-177 PSMA-617, organ retention can be reasonably approximated by a mono-exponential or bi-exponential function over the 0-144 hour post-injection window.
The fitting itself is not computationally expensive. What matters is how you handle the early time points. Biodistribution in the first few hours after injection is driven by both specific and non-specific binding, and for some organs the time-activity curve will peak before the first scheduled imaging time point. A fitting routine that assumes zero activity at time zero and builds upward from the first measured point may underestimate total cumulated activity by 15 to 30% for organs with early peaks.
At YSOTOPE, we address this with an empirical uptake model that adds a parameterized uptake phase prior to the first measured point, with priors drawn from published population data for each radiopharmaceutical-organ pair. The population prior helps constrain the fit when early imaging data is unavailable, without overriding the patient's own measurements when early points exist.
Dose Calculation: S-Values vs. Voxel-Based Kernels
The final dose calculation step has two main approaches. OLINDA/EXM-style S-value dosimetry uses pre-tabulated absorbed fractions for standardized phantom geometries. It is fast and well-validated, but it does not account for patient-specific organ size and shape. A kidney that is 30% smaller than the standard phantom kidney will have a different self-absorbed dose for the same cumulated activity, because the geometric relationship between emission source and absorbing tissue is different.
Voxel-based dose kernel convolution accounts for patient geometry but requires more memory and compute. For a single 256x256x80 voxel SPECT volume, the convolution is manageable in seconds on modern hardware. For a full Monte Carlo dose calculation, the computational demand increases by orders of magnitude, and the marginal accuracy gain over a well-validated point-kernel method is rarely justified in routine clinical dosimetry.
We use a patient-scaling correction on top of MIRD S-values as a practical middle ground. The organ-specific cumulated activity is scaled by the ratio of patient organ volume to reference phantom volume before S-value lookup. This captures the dominant geometric effect without the full computational cost of voxel convolution, and the correction is transparent enough to validate against manual checks.
Reporting That Clinicians Can Act On
The output of a dosimetry platform is a number in gray. But a number in gray with no clinical context is not a decision. The dose report needs to communicate: absorbed dose to each structure, the dose constraint for each organ at risk, the recommended activity for the next cycle relative to the current cycle, and a plain-language interpretation of whether the patient is within a safe and effective dose range.
We found that the format of the report mattered as much as its content. A dense table of organ doses with no reference values embedded required clinicians to maintain their own lookup table for dose thresholds. Building the reference values into the report, with clear threshold lines and cycle-to-cycle trend tracking, reduced the cognitive load of clinical decision-making. More importantly, it reduced the chance of a busy clinician misinterpreting a report under time pressure.
Where the Problem Is Not Yet Solved
Multi-lesion dosimetry for patients with many small metastases remains technically hard. Partial volume effect in SPECT becomes significant when lesion diameter falls below approximately 2 cm, and many NET and prostate cancer patients have sub-2 cm liver metastases that are clinically relevant but not reliably quantifiable by standard SPECT. Better spatial resolution from newer SPECT/CT systems helps but does not eliminate this constraint.
We consider partial volume correction an active area and note the methods honestly in our reporting when lesion sizes approach the SPECT resolution limit. A platform that does not document this limitation is not being rigorous. It is obscuring uncertainty that clinical teams need to account for when they interpret dose numbers and adjust treatment plans.
The architectural work of connecting all these steps into a clinical-grade pipeline is ongoing. Each center that joins our program adds real-world edge cases that sharpen the pipeline. That is the nature of building infrastructure for a technically demanding clinical problem, and we consider it part of the job rather than a limitation of the approach.