médical IA

Prostate cancer: segmenting lesions on PSMA PET/CT with a transformer, and stratifying survival before radioligand therapy

Published on June 22, 2026 · 9 min read

Fine-UNETR is a vision transformer that automatically segments PSMA-avid lesions on whole-body prostate-cancer PET/CT, then extracts tumor-burden biomarkers from them. Trained on 299 scans and tested on 74, it reaches a Dice coefficient of 66.63% internally, which drops to 44.11% on an external dataset — while still preserving detection of the most intense lesions. The derived biomarkers (total tumor volume, total uptake) track ground truth almost perfectly and significantly stratify the overall survival of 67 patients before radioligand therapy. A credible automation of quantification, but one whose fine segmentation remains fragile outside its home center.

The context

Advanced prostate cancer is increasingly managed along a theranostic logic: the same molecular target serves both diagnosis and treatment. That target is PSMA (prostate-specific membrane antigen), a protein overexpressed on the surface of tumor cells. In imaging, a radioactive tracer that binds to it enables PSMA PET/CT — positron emission tomography coupled with CT — which lights up metastases as bright spots across the whole body. In therapy, the same vector, now loaded with a radiation-emitting isotope, delivers a dose directly onto the lesions: this is lutetium-177 PSMA-617 radioligand therapy, which became a standard for metastatic castration-resistant prostate cancer after the VISION trial.

To select patients, monitor response and estimate prognosis, physicians need to measure total tumor burden: how many lesions, what volume, what uptake intensity. But counting and delineating dozens of lesions by hand on a whole-body scan is slow, tedious and operator-dependent. Hence the idea of automating this segmentation with deep learning. That is exactly what Fine-UNETR proposes.

The method

The authors start from the UNETR architecture, a segmentation network that pairs a vision transformer (the attention mechanism popularized by large language models, here applied to 3D image blocks) with a classic U-shaped decoder to reconstruct the lesion mask. Their variant, Fine-UNETR, refines the encoding by splitting the volume into very small 8×8×8-voxel blocks — hence the "Fine" — and trains with a sliding window along the body axis, to handle whole-body images that would not fit in memory at once.

The dataset is a retrospective cohort of 373 PSMA PET/CT scans from prostate-cancer patients (mean age 71 ± 8 years), split into 299 training and 74 test scans. To measure generalization, the authors add two distinct evaluations. First an external validation on 192 cases from the public AutoPET IV dataset (an international PET/CT lesion-segmentation challenge) — that is, images acquired elsewhere, on other machines, with other protocols. Second, an independent cohort of 67 patients scanned before radioligand therapy, to test the prognostic utility of the automatic biomarkers on overall survival, analyzed with Kaplan-Meier curves and the log-rank test.

Four families of measures judge the model. For delineation quality: the Dice coefficient (DSC), which quantifies the overlap between the predicted mask and the reference mask (100% = perfect overlap), plus sensitivity and precision. For detection: the rate of correctly spotted lesions. For quantification: the correlation between AI-computed biomarkers and ground truth (total tumor volume, total uptake, lesion count). And for the clinic: the ability of these biomarkers to separate patients by survival.

The results

On the internal test, Fine-UNETR achieves a Dice of 66.63%, a sensitivity of 70.27% and a precision of 67.77%. The lesion detection rate is 79.53% across all lesions, but climbs to 96.05% if you count only the most intense ones (SUVmax ≥ 5, SUV being the standardized measure of tracer uptake). In other words, the model sees lesions that "shine" brightly very well, and the fainter ones markedly less well.

On the AutoPET IV external validation, the Dice collapses to 44.11%: voxel-by-voxel delineation becomes unreliable on images from elsewhere. Yet the lesion detection rate stays high, at 87.18%. The model therefore keeps spotting tumor foci, even when it traces their outlines poorly.

It is at the patient level that the figures become reassuring: the agreement between AI biomarkers and ground truth is near-perfect — correlation of 0.984 for total tumor volume, 0.989 for total uptake, 0.960 for lesion count. A correlation of 0.98 means that, even if the exact contours differ, the overall tumor-burden estimate faithfully follows the reference measure. Finally, in the pre-therapy cohort of 67 patients, three automatic biomarkers significantly separate survival curves: total tumor volume (p = 0.0019), SUVmax (p = 0.014) and SUVmean (p = 0.016). Concretely: patients the AI ranks as having a high tumor burden die earlier, confirming that the automatic measure captures real prognostic information.

What is good

The detection / delineation split, owned and measured. Rather than hiding the weak external Dice, the authors show that lesion detection holds (87.18%) where voxel overlap drops (44.11%). This is a clinically relevant distinction: to estimate a tumor burden, knowing where the lesions are often matters more than tracing their boundary to the voxel.

A genuine external validation, on an independent dataset. Many imaging papers settle for an internal test. Here, AutoPET IV (192 cases) provides a trial on images acquired outside the home center — the most honest test of generalization, and the one that precisely reveals the performance drop.

The loop closed all the way to prognosis. The study does not stop at a technical metric: it links the automatic biomarkers to overall survival in an independent pre-radioligand-therapy cohort. This is exactly the clinical translation missing from most segmentation work, and it gives concrete meaning to the machine-measured tumor volume.

What is less good

Fragile generalization: the domain-shift failure mode. The Dice going from 66.63% to 44.11% between internal and external is the hands-on illustration of population and protocol bias (domain shift): a model trained on the machines and patients of a single center sees its contours degrade elsewhere. For any use that relies on volume accuracy (for example estimating a dose), a delineation with 44% overlap is insufficient as it stands.

Detection that flatters on small lesions. The jump from 79.53% to 96.05% detection depending on whether faint lesions are included is a textbook case of a misleading metric: the flattering performance is carried by the easy, bright lesions. Yet it is often the small, discreet, low-uptake lesions that make the difference for early detection or a complete staging workup — precisely the ones the model misses most.

An absent comparator, and transparency unknowns. The abstract reports no comparison to an established segmentation reference such as nnU-Net: it is impossible to tell whether Fine-UNETR does better than the state of the art or merely as well. Add the expected limits — a retrospective study, a small survival cohort (67 patients) without its own external validation, high tumor burden long known as a poor-prognosis factor (so the AI automates a useful measure more than it discovers a signal), and an abstract silent on code availability, funding and conflicts of interest. Finally, coupling to the actual dosimetry of radioligand therapy, which would be the most valuable outlet, is not evaluated.

What it changes

For the research community, the message is nuanced and useful: on PSMA PET/CT, chasing a perfect Dice may not be the right goal. If the aim is to quantify whole-body tumor burden, the robustness of detection and the fidelity of aggregated biomarkers matter more than voxel-by-voxel overlap — a reframing of evaluation metrics that other oncology segmentation work would do well to adopt.

For clinicians, the immediate contribution is potential, not settled: a tool that pre-segments lesions and automatically computes a survival-correlated tumor volume would save considerable time in nuclear medicine, provided prospective, multicenter validation and a comparison to a standard. As it stands, the external fragility forbids relying on the contours for a fine quantitative decision. For patients, the stake is concrete: measuring tumor burden better means better selecting those who will benefit from radioligand therapy and tracking their response more closely — but a prognostic measure is not a prescription, and it is the medical team that interprets the number in context.