Using Time as the Label: Click-Supervised Delivery Maps for Radiolabeled siRNA and ASO Imaging

Drug developers can now attach a radioactive tag to an RNA-interference drug and watch where it goes in a living body with a nuclear-medicine scanner. The drugs in question are small interfering RNA (siRNA) or antisense oligonucleotides (ASOs), short synthetic nucleic acids that silence a target gene. The scanners are PET or SPECT, which image the tagged drug's radioactivity in 3D alongside a CT scan of anatomy. The bottleneck is turning those scans into numbers. A pharmacologist wants the drug's concentration in the tumor, or in the spinal cord, at each time point. That means an expert must draw a region of interest (ROI) on every scan of every subject at every time point. The images make this hard. Kidneys and liver clear these molecules and glow far brighter than the target, and SPECT blur smears that glow into neighboring tissue.

This note proposes to replace most of that drawing with two sources of supervision the data already contains. The first is automatic organ masks computed from the CT. The second, and the main one, is the fact that each region has its own activity curve over time: clearance organs fill and drain, while a target that takes up the drug retains it. The expert supplies one click per target on one time point. A small 3D network then learns a map of where the target is, trained so that organ-shaped kinetic components, blurred by the scanner, plus one target component, explain the whole dynamic series. The claim is that this gives target time-activity curves as accurate as expert ROIs at a small fraction of the annotation cost. The gain should be largest exactly where current tools fail, which is targets sitting next to the kidneys or liver.

What the field has and what it lacks

On the imaging-chemistry side, the tracers already exist, but analysis is manual. Noninvasive Visualization of RNA Delivery with 99mTc-Radiolabeled Small-Interference RNA in Tumor Xenografts labeled an anti-hTERT siRNA with technetium-99m and reported tumor-to-blood and tumor-to-muscle ratios of 2.62 and 6.02 at 6 hours, alongside intense renal and hepatic uptake. Central Nervous System Biodistribution and Pharmacokinetics of Radiolabeled Tofersen in Rodents, Nonhuman Primates, and Humans imaged a technetium-labeled ASO by SPECT/CT in three species. It found that kinetics differ across species: human brain uptake rose over the first 4 hours while it fell in rodents and primates. It also noted that SPECT resolution limits separating fluid spaces from tissue. These papers motivate the proposal in three ways. The quantities of interest are curves over time rather than single masks. Clutter from clearance organs dominates the image. And any learned "normal pattern" is unlikely to transfer across species or across new drugs.

The machine-learning work closest to this problem is on diagnostic PET, and it constrains the design. Disentangled PET Lesion Segmentation separates normal bladder and kidney uptake from lesions in FDG PET. It uses a latent space that an adversarial critic forces to match a healthy-subject population, and it reported Dice 0.656 against 0.057 for a plain 3D U-Net on the lower torso. Two things block its use for a new RNAi drug: it needs a cohort of healthy scans for that same tracer, and it needs full lesion masks. The autoPET III entries From FDG to PSMA and Kalisch et al. show that adding automatic TotalSegmentator organ labels as an auxiliary task cuts false positives in excretion organs. LesionTracer reached global Dice 68.40 versus 57.61 for default nnU-Net, with false-positive volume falling from 19.32 to 7.82 ml. Both entries still train on hundreds of fully annotated scans at a single time point.

We borrow their organ-label trick but use it differently. The organ masks define where "known-kinetics" components may live. They are not an auxiliary prediction target.

On the annotation side, Weakly Supervised Lesion Co-Segmentation on Ct Scans trains from routine clinical diameter measurements (RECIST lines) and reached Dice 90.3% on DeepLesion. Learning to Segment Medical Images with Scribble-Supervision Alone trains from scribbles and loses only 2.9 to 4.5 Dice points on MRI. Both rely on sharp intensity edges, via GrabCut or random-walk seeding and iterative CRF refinement. SPECT emission images lack such edges. The two papers also disagree on CRF refinement, since Agarwal et al. found that dense CRFs hurt recall. That warns against importing edge-based pseudo-labeling wholesale. The Role of Pleura and Adipose in Lung Ultrasound AI shows that unmasked models latch onto irrelevant anatomy. This supports restricting what each component of our model may explain.

Core Claim

In multi-time-point imaging of a radiolabeled macromolecule, most voxels' activity over time can be explained by a handful of organ-level curves, each confined to a CT-derived organ mask and blurred by the scanner's known point-spread function. Whatever remains, near a single expert click, is the target. Modeling this explicitly supplies the "healthy" component that PET-Disentangler had to learn from a healthy population. It does so within each subject, so it needs no healthy cohort and no cross-species training.

Concretely, let $A(x,t)$ be measured activity at voxel $x$ and time $t$, corrected for radioactive decay. We model

$$A(x,t) \approx \Big(h * \Big[\textstyle\sum_{k} m_k\, c_k(t) + f\, c_T(t) + b\, c_B(t)\Big]\Big)(x),$$

where $m_k$ is the TotalSegmentator mask for organ $k$ (kidneys, liver, bladder, spleen, spinal canal and so on), $c_k(t)$ is that organ's curve, $f(x)\in[0,1]$ is the target map predicted by a 3D network from the CT and all time frames, $c_T(t)$ is the target curve, and $b$ with curve $c_B$ is a diffuse blood and soft-tissue background. The symbol $h$ is the scanner point-spread function and $*$ denotes convolution.

The network and the curves are fit per subject by maximizing a Poisson likelihood of the raw counts. The loss adds a seed term forcing $f=1$ at the clicked voxel and a small neighborhood, a sparsity and smoothness penalty on $f$, and a penalty on $f$ inside organ-mask cores:

$$\mathcal{L} = -\log p_{\text{Poisson}}(A \mid \theta) + \lambda_s \|f\|_1 + \lambda_{tv}\,\mathrm{TV}(f) - \lambda_c \log f(x_{\text{click}}).$$

Here $\theta$ collects the network weights and all curves, $\mathrm{TV}$ is total variation, and the $\lambda$'s are weights. The network is shared across subjects so that it learns what a target looks like in CT and emission. The curves are always per subject.

Research Program

The work has three parts. The first is a simulation benchmark with exact ground truth, since no public macromolecule SPECT dataset is available to us. We take autoPET CT volumes with their expert lesion masks (The autoPET challenge) and their TotalSegmentator organ masks. We treat the lesions as delivery targets and assign compartment-model curves to organs and targets. The curve shapes will be set by hand to match the qualitative patterns reported for the siRNA and tofersen studies: early renal peak with washout, slow hepatic accumulation, and target retention. We then blur to SPECT resolution, decay-correct technetium's roughly 6-hour half-life, and add Poisson noise at realistic counts for 4 to 6 frames over 24 hours. We will explicitly vary target distance from the kidney and liver, target size relative to scanner blur, and how similar the target curve is to its neighbor's curve.

The second part builds the model above, using a lightweight 3D U-Net backbone. The third part is real-data validation on a preclinical technetium-siRNA or ASO xenograft study obtained through a collaborating imaging lab. That study would follow the protocol of Kang et al., with expert-drawn ROIs at every time point on a held-out set of animals. If the collaborator can share multi-species CNS SPECT, we will also test cross-species behavior. Access to such data is the main practical risk and is stated as a precondition.

Baselines

The naive reference is the current practice approximated automatically. We take the TotalSegmentator organ mask for organs, and for the target, the expert's click plus a fixed-radius sphere or a 41%-of-maximum threshold grown from the click. Each is read out at every time point on the same frames. The upper reference is the expert's own ROIs drawn at every time point. Beating the naive readout means lower error in the target's area under the curve and in target-to-blood ratio. The error must be lower overall and specifically for targets within two blur-widths of the kidney or liver.

The strongest learned competitor is the LesionTracer recipe from From FDG to PSMA: nnU-Net with a residual encoder and organ multitask supervision, which reported global Dice 68.40 on autoPET III. We will run it two ways. One is pretrained on autoPET and fine-tuned per time point. The other gets full masks on $K$ subjects, where $K$ is chosen so that its expert annotation time equals the click budget our method uses. Both runs use the same frames, GPU budget, and hyperparameter search size. Beating it means lower curve error at equal expert time. Matching it at a tenth of the expert time also counts.

The closest structural competitor is Disentangled PET Lesion Segmentation, which reported Dice 0.656 on lower-torso FDG. We will run it per time point with its healthy critic trained on simulated target-free scans, which is an advantage it would not have for a real new drug. On autoPET simulations we additionally report Dice against the known masks.

Classical factor analysis or kinetic clustering of dynamic images is the non-deep competitor for voxel-curve unmixing. We found no reported result for it on Lacuna. We will implement a standard nonnegative factorization of the voxel-by-time matrix, seeded at the click, with an equal tuning budget.

Three ablations isolate the key ingredient. The first removes time by fitting on the single best frame. The second removes the organ masks by letting components live anywhere. The third removes the point-spread-function term. If the single-frame ablation does as well as the full model, the central claim is wrong.

Why not just…?

Why not just read out TotalSegmentator organ masks and let the expert draw the target? For organs this is nearly the right answer, and we adopt it inside the model rather than compete with it. It fails in two places. SPECT blur makes a CT-exact mask either miss activity at the edges or absorb spill-over from neighbors. And targets such as tumors or subregions of the neural axis have no anatomical mask, so the expert must still draw them at every frame; TotalSegmentator-style tools do not segment lesions and are already weak on small structures, with adrenal Dice around 0.56 reported in MRSegmentator. The naive-readout baseline measures exactly how much the kinetic model adds. If it adds nothing far from clearance organs, we will say the method is only worth using near them.

Why not just threshold the late frame, after the kidneys have cleared? Late technetium frames are count-starved because of the short half-life, and hepatic retention persists. The 3D U-Net collapse to Dice 0.057 reported by Disentangled PET Lesion Segmentation shows that intensity alone confuses excretion with lesions. The simulation varies count level, so the threshold baseline will show where it breaks.

Why not just fine-tune a supervised PET/CT segmenter on a few annotated animals? This is the real threat, since autoPET-pretrained nnU-Nets are strong and data curation often matters more than architecture (The autoPET challenge). Two concerns remain. Kalisch et al. found that tracer-specific biodistribution defeats universal models. And a new drug in a new species offers no labeled pool large enough for this approach. The equal-expert-time comparison settles the question directly. If fine-tuning wins, the contribution reduces to the simulation benchmark and the per-subject kinetic prior becomes a regularizer, not a replacement.

Proposed Experiments

On the simulation benchmark, the primary metric is relative error of the target area under the curve and of the target-to-blood ratio against ground truth. It will be reported as a function of distance to the nearest clearance organ and of target size. Secondary metrics are Dice against the known lesion mask and expert seconds per subject.

The decisive check has two parts. First, near-organ targets must show at least halved curve error relative to the naive readout. Second, the single-frame ablation must lose most of that gain. Together these establish that time, and not the network, does the work.

On real xenograft data, we compare each method's curves and 6-hour tumor-to-blood and tumor-to-muscle ratios against the expert's full ROI series. For context we use the magnitudes that Kang et al. reported (2.62 and 6.02), but we do not use them as test labels. If multi-species CNS data are available, a final check fits the model per subject in rats, primates, and humans with no retraining of the curves. It then tests whether the reversed brain-uptake trend between humans and other species, reported for tofersen, is recovered. A learned population prior would be expected to miss this reversal. A within-subject prior should not.