autoPET IV: Interactive Lesion Segmentation

• autoPET IV is a benchmark that redefines PET/CT lesion segmentation by framing it as a human-in-the-loop task with interactive prompts and multi-tracer, multi-center data.
• It implements innovative methods such as prompt encoding using EDT maps, tracer-aware pipelines, and organ supervision to enhance segmentation accuracy and reduce false positives.
• Verified longitudinal tracking leverages clinician-verified prompts to condition lesion delineation on baseline appearances, significantly improving DSC performance.

autoPET IV is a challenge setting within the autoPET series that frames lesion analysis in PET/CT as a human-in-the-loop problem rather than a purely automatic segmentation task. In the associated literature, it is characterized by interactive segmentation with simulated user prompts, multi-tracer and multi-center PET/CT data spanning FDG and PSMA, and, in challenge-winning work, clinician-verified interactive lesion tracking that conditions follow-up delineation on both a verified prompt and the baseline lesion appearance (Huang et al., 2 Sep 2025, Rokuss et al., 29 Aug 2025, Kirchhoff et al., 22 May 2026).

1. Position within the autoPET research program

The autoPET line of work emerged from whole-body lesion segmentation in FDG PET/CT. AutoPET Challenge 2022 addressed “automatic segmentation of whole-body tumor lesion based on deep learning and FDG PET/CT,” and contemporary submissions emphasized systemic whole-body modeling under substantial physiological uptake confounds (Zhong et al., 2022). Another 2022 contribution described the challenge as providing a public data set with FDG-PET/CT scans from 900 patients, and its solution combined nn-Unet, Swin UNETR, and a maximum intensity projection classifier to reduce false positives in healthy cases (Heiliger et al., 2022).

By 2024, the benchmark had expanded to “generalizable lesion segmentation in multi-tracer whole-body PET-CT imaging.” The reported dataset comprised 1611 co-registered PET-CT volumes, with an FDG cohort of 1014 subjects and a PSMA cohort of 597 subjects, and the task was accurate binary segmentation of tracer-avid lesions in either FDG or PSMA PET-CT scans (Astaraki et al., 2024). autoPET IV builds directly on that trajectory: one associated paper states that, “Building on this foundation, autoPET IV introduces a human-in-the-loop scenario to efficiently utilize interactive human guidance in segmentation tasks” (Huang et al., 2 Sep 2025).

This evolution is methodologically significant. Earlier autoPET editions centered on fully automatic whole-body delineation; autoPET IV shifts the benchmark toward workflows in which model predictions are explicitly updated by human guidance, while retaining the series’ emphasis on difficult physiological uptake patterns, cross-center heterogeneity, and lesion variability.

2. Benchmark formulation, data regime, and evaluation

The autoPET IV setting uses a multi-tracer, multi-center PET/CT corpus. One submission specifies a dataset of 1,014 FDG and 597 PSMA PET/CT scans, while another describes the main challenge data as combining FDG-PET/CT and PSMA-PET/CT multi-tracer, multi-center datasets (Huang et al., 2 Sep 2025, Rokuss et al., 29 Aug 2025). In the 2024 multi-tracer formulation, the FDG cohort contained 501 subjects with histologically confirmed cancer and 513 controls, whereas the PSMA cohort contained 537 prostate carcinoma subjects and 60 controls (Astaraki et al., 2024).

The core segmentation target is binary lesion masking. A 2024 challenge contribution states that pathological regions were identified via tracer avidity and clinical review, with manual delineation of tumor lesions on axial slices, and formalizes the task as accurate binary segmentation of tracer-avid lesions (Astaraki et al., 2024). autoPET IV extends this by associating each scan with interactive guidance. In one integrated pipeline, each scan receives up to 10 simulated foreground and 10 background clicks, represented as 3D Gaussian maps added to the model input, yielding four channels: CT, PET, foreground clicks, and background clicks (Huang et al., 2 Sep 2025).

The standard reported metrics are the Dice coefficient, false positive volume, and false negative volume. One autoPET IV submission additionally reports an AUC calculated over 0, 3, 7, and 10 clicks (Rokuss et al., 29 Aug 2025). The Dice Similarity Coefficient is written as

$\mathrm{DSC} = \frac{2|X \cap Y|}{|X| + |Y|}$

with $X$ denoting predicted lesion voxels and $Y$ denoting ground-truth lesion voxels (Huang et al., 2 Sep 2025). In the challenge ecosystem, false positive volume measures predicted lesion tissue not supported by ground truth, and false negative volume measures missed lesion burden.

The presence of click-conditioned evaluation distinguishes autoPET IV from earlier autoPET editions. Performance is no longer summarized solely by zero-guidance segmentation quality; instead, papers report how Dice, FPV, and FNV evolve as interaction count increases.

3. Promptable segmentation and simulated interaction

A central autoPET IV design axis is the conversion of sparse user input into network-consumable spatial signals. A leading promptable submission extends a winning autoPET III nnU-Net pipeline by adding two extra input channels, one for foreground user prompts and one for background user prompts, and systematically compares two encoding families: 3D Gaussian kernels and Euclidean Distance Transform (EDT) maps (Rokuss et al., 29 Aug 2025). Gaussian clicks are centered at the prompt with various standard deviations, whereas EDT channels store voxelwise Euclidean distance to the prompt location.

In that study, EDT encodings consistently outperformed Gaussian kernels in cross-validation, particularly for Dice scores and false positive volume, and were therefore chosen for the final model (Rokuss et al., 29 Aug 2025). This directly counters the common assumption that prompt representation is a secondary implementation detail. Within autoPET IV, prompt encoding is itself a major modeling variable.

Training-time interaction simulation is similarly explicit. The promptable nnU-Net system performs online simulation of user clicks after data augmentation and within the sampled training patches. Between 0 and 10 foreground and 0 and 10 background prompts are simulated per patch, with a logarithmic distribution favoring fewer points. Validation uses only the official pre-computed prompts for fair comparison. To improve robustness to realistic prompting behavior, the same work introduces a custom 20% sampling strategy in which points are sampled less restrictively, anywhere inside the target region with a bias toward the lesion core (Rokuss et al., 29 Aug 2025).

A separate integrated autoPET IV pipeline studies two alternative interaction regimes: full-guidance training, which always includes all 10 foreground/background clicks, and stochastic click sampling, which randomly selects 0–10 foreground/background clicks per sample (Huang et al., 2 Sep 2025). It then deploys a hybrid inference rule: for sparse clicks, 0–4, the stochastic-trained model is used; for dense clicks, at least 5, the full-guidance-trained model is used (Huang et al., 2 Sep 2025). This suggests that autoPET IV is not merely a benchmark for prompt-conditioned inference, but also a benchmark for matching training distributions to anticipated clinical interaction density.

4. Tracer-aware and organ-aware pipeline design

autoPET IV submissions frequently integrate tracer classification and anatomical supervision upstream of lesion delineation. One pipeline implements tracer classification as a preprocessing step using a pretrained model with two ResNets on coronal and sagittal MIP views, with features combined and fed into an MLP for FDG/PSMA discrimination. The reported accuracy is 100% on the dataset (Huang et al., 2 Sep 2025). This enables tracer-specific downstream processing rather than forcing a single lesion model to absorb all tracer heterogeneity.

The same work augments the segmentation network with organ supervision through a dual-head architecture. The additional head predicts 10 anatomical structures, using pseudo-annotations from TotalSegmentator, and the loss is equally weighted between lesion and organ segmentation heads (Huang et al., 2 Sep 2025). In that formulation, organ supervision is introduced to enhance anatomical awareness and reduce false positives. Post-processing remains tracer-specific as well: predictions with standardized uptake value below 1.5 for FDG or 1.0 for PSMA are removed (Huang et al., 2 Sep 2025).

Promptable models in autoPET IV also inherit strong pretraining and augmentation protocols from the broader nnU-Net ecosystem. One ensemble uses the ResEncL U-Net preset, large-scale multi-modal pretraining via MultiTalent, organ supervision to reduce false positives, and heavy data augmentation including PET–CT misalignment simulation (Rokuss et al., 29 Aug 2025). These details indicate that autoPET IV methods are not defined by interaction channels alone; they combine click conditioning with tracer recognition, anatomical priors, and aggressive robustness engineering.

A plausible implication is that autoPET IV operationalizes lesion segmentation as a structured decision pipeline rather than a monolithic voxel classifier. Tracer identity, organ context, prompt density, and post-hoc SUV constraints are treated as first-class variables.

5. Verified longitudinal tracking

A distinct and technically important autoPET IV-associated direction is clinician-verified interactive lesion tracking. In “Verified Tracking,” a registration network proposes the most likely follow-up lesion center by mapping the baseline position $p_0$ through a deformation field $\phi : I_0 \to I_t$,

$\hat{p}_t = \phi(p_0),$

after which a clinician verifies or corrects $\hat{p}_t$ to obtain $p_t$. Segmentation is then conditioned on both the corrected prompt and the baseline lesion (Kirchhoff et al., 22 May 2026).

The corresponding network uses early spatial prompt fusion and latent temporal difference weighting. Baseline and follow-up inputs are encoded as

$X_0 = [I_0, G(p_0)], \qquad X_t = [I_t, G(p_t)],$

where $G(p)$ is a Gaussian heatmap channel at the prompt location. At each U-Net skip level $X$0, temporal difference weighting combines multiscale features as

$X$1

The stated goal is to force the network to focus on longitudinal changes rather than ignore the baseline branch (Kirchhoff et al., 22 May 2026).

The ablation results are particularly informative. A single-timepoint promptable baseline achieved 55.8 DSC, while naive longitudinal concatenation without synthetic pretraining dropped to 45.8 DSC. Adding prompt simulation raised naive concatenation to 54.0, and combining prompt simulation, synthetic pretraining, and difference weighting reached 58.5 (Kirchhoff et al., 22 May 2026). This addresses another recurrent misconception: longitudinal context is not automatically beneficial. In the reported autoPET IV setting, exploiting prior scans required explicit temporal fusion and large-scale synthetic pretraining.

This work reports 60.7 DSC in the automatic paradigm and 73.7 DSC in the verified paradigm on autoPET IV, and states that it “secured first place in the MICCAI autoPET IV challenge” (Kirchhoff et al., 22 May 2026). The verified formulation is conceptually aligned with the general autoPET IV emphasis on safe human oversight: the prompt is not an optional refinement after segmentation, but a verified interaction that eliminates silent retrieval failures.

6. Reported results and broader methodological significance

The autoPET IV literature reports a consistent benefit from interactive guidance, but the gain depends strongly on prompt encoding, supervision strategy, and whether the task is single-timepoint segmentation or longitudinal tracking.

Method family	Core idea	Reported result
Promptable nnU-Net ensemble	EDT-encoded foreground/background prompts	76.35 Dice, 8.50 FPvol, 3.76 FNvol, AUC 2.22
Tracer-aware organ-supervised pipeline	Tracer classification, organ supervision, simulated clicks	0-click DICE 0.788 and 10-click DICE 0.889
Verified longitudinal tracking	Early prompt fusion and latent temporal difference weighting	73.7 DSC in verified tracking

In the promptable nnU-Net submission, the strongest cross-validation setting was an ensemble of all 5 folds of two EDT-based models, one using a custom point strategy and one adding external data; it achieved Dice 76.35, FPvol 8.50, FNvol 3.76, and AUC 2.22. Without external data, the EDT Size 2 model with custom points achieved Dice 76.19, FPvol 8.05, FNvol 4.19, and AUC 2.21 (Rokuss et al., 29 Aug 2025). In the integrated organ-supervised pipeline, the PSMA-specific V2 model improved from DICE 0.619 at 0 clicks to 0.871 at 10 clicks, while FPV decreased from 0.991 to 0.332 and FNV decreased from 5.641 to 1.613. In the unified FDG+PSMA setting, V4 reported DICE 0.788 at 0 clicks, and V3 reported DICE 0.889 at 10 clicks (Huang et al., 2 Sep 2025).

These results are best understood against the preceding autoPET literature. Before autoPET IV, fully automatic multi-tracer segmentation on the AutoPET 2024 dataset reported a mean Dice Similarity Coefficient of 0.6687, mean false negative volume of 10.9522 ml, and mean false positive volume of 2.9684 ml on the preliminary test phase for Task-1 using a 3D Residual UNet with Generalized Dice Focal Loss (Ahamed, 2024). Earlier interactive annotation research had already shown that full-volume click-guided lesion delineation was practical: SW-FastEdit, operating on complete PET volumes with a sliding window-based interactive scheme, reached Dice@10 of 85.50% on the AutoPET validation set and reported annotation times of 6–8 min per volume with a low perceived NASA-TLX workload (Hadlich et al., 2023). autoPET IV can therefore be read as a formalization of a transition already visible in the ecosystem: from benchmark segmentation toward benchmark interaction.

A broader methodological consequence is that data-centric and interaction-centric strategies are converging. On the AutoPET dataset, an adapted DiffTumor pipeline generated three synthesized PET-CT samples from each of the 1,291 training cases, producing an augmented training set of 5,152 cases, and the resulting segmentation model improved from 0.3650 to 0.4542 Dice under one transform and from 0.5859 to 0.6179 Dice under 30 transforms (Chan et al., 2024). This suggests that future autoPET IV-style systems may combine promptability, organ supervision, tracer awareness, and synthetic augmentation rather than treat them as separate research threads.

In that sense, autoPET IV is less a single architecture than a benchmark philosophy: lesion segmentation in PET/CT is evaluated under realistic heterogeneity, explicit human guidance, and, in its most advanced formulation, verified longitudinal follow-up.