Paper deep dive
BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net
Marven Sherif, Amgad Elmasry, Youssef Ghazal, Ayman Elghotni
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 9/8/2026, 3:30:52 AM
Summary
This paper presents an interactive lesion segmentation method for whole-body PET/CT scans, submitted to the autoPET/CT V challenge. The approach uses a scribble-conditioned residual encoder U-Net that takes CT, PET, and user-provided foreground/background scribbles as inputs. The model is initialized from the autoPET-III winning weights and fine-tuned using a normalization strategy based on aorta blood-pool activity. Results show that interactive correction significantly improves segmentation performance, with the first scribble accounting for 85% of the gain, raising mean Dice from 0.554 to 0.751 after five rounds.
Entities (11)
Relation Signals (9)
BS: Take the Hint → uses → Scribble-Conditioned ResEnc U-Net
confidence 99% · We present our submission, a scribble-conditioned residual encoder U-Net
BS: Take the Hint → achievesmetric → Dice
confidence 98% · mean Dice is 0.554... rising to 0.751... after five correction rounds
BS: Take the Hint → achievesmetric → F1
confidence 98% · mean lesion-level F1 is 0.528... rising to 0.733... after five correction rounds
BS: Take the Hint → participatedin → autoPET/CT V
confidence 97% · The autoPET/CT V challenge addresses this by making segmentation interactive... We present our submission
Scribble-Conditioned ResEnc U-Net → initializedfrom → autoPET-III
confidence 95% · The network is initialised from the autoPET-III winning weights
autoPET/CT V → usestracers → PSMA
confidence 95% · the cohort contains both 18F-FDG and 68Ga/18F-PSMA studies
autoPET/CT V → usestracers → 18F-FDG
confidence 95% · the cohort contains both 18F-FDG and 68Ga/18F-PSMA studies
BS: Take the Hint → developedby → Brightskies
confidence 92% · Affiliation: Brightskies, Egypt... We present our submission
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Automated lesion segmentation in whole-body PET/CT is complicated by the variety of physiological tracer uptake patterns and by the differing appearance of lesions across tracers. The autoPET/CT V challenge addresses this by making segmentation interactive: user scribbles marking foreground and background are supplied alongside the image, and the algorithm is expected to exploit them. We present our submission, a scribble-conditioned residual encoder U-Net operating on four input channels: CT, PET, and a sparse scribble map for each of foreground and background. The network is initialised from the autoPET-III winning weights and extended from two to four input channels, with the two scribble channels zero-initialised so that the pretrained representation is preserved exactly at initialisation. Every model is fine-tuned per fold from the corresponding autoPET-III fold checkpoint, so that no validation case is seen during pretraining. PET intensities are normalised against a per-scan aorta blood-pool reference derived from a CT segmentation, which removes tracer- and centre-specific scaling without requiring lesion labels. At inference the five fold models are ensembled by averaging their softmax outputs per sliding-window patch, before Gaussian-weighted stitching. On the challenge's five-fold split, with each fold evaluated on its own validation cases, mean Dice is 0.554 and mean lesion-level F1 is 0.528 without scribbles, rising to 0.751 and 0.733 after five correction rounds. About 85% of that gain follows the first scribble, and the spread between fold models narrows five-fold over the same rounds, so interaction largely compensates for how well or badly a given model segments unaided.
Tags
Links
- Source: https://arxiv.org/abs/2609.01554v1
- Canonical: https://arxiv.org/abs/2609.01554v1
Trouble viewing inline? Open PDF directly →
Full Text
32,299 characters extracted from source content.
Expand or collapse full text
BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net Marven Sherif Affiliation: Brightskies, Egypt E-mail marven.sherif,amgad.elmasry,youssef.ghazal,ayman.elghotni@brightskiesinc.com Amgad Elmasry Affiliation: Brightskies, Egypt E-mail marven.sherif,amgad.elmasry,youssef.ghazal,ayman.elghotni@brightskiesinc.com Youssef Ghazal Affiliation: Brightskies, Egypt E-mail marven.sherif,amgad.elmasry,youssef.ghazal,ayman.elghotni@brightskiesinc.com Ayman Elghotni Affiliation: Brightskies, Egypt E-mail marven.sherif,amgad.elmasry,youssef.ghazal,ayman.elghotni@brightskiesinc.com Abstract Automated lesion segmentation in whole-body PET/CT is complicated by the variety of physiological tracer uptake patterns and by the differing appearance of lesions across tracers. The autoPET/CT V challenge addresses this by making segmentation interactive: user scribbles marking foreground and background are supplied alongside the image, and the algorithm is expected to exploit them. We present our submission, a scribble-conditioned residual encoder U-Net operating on four input channels – CT, PET, and a sparse scribble map for each of foreground and background. The network is initialised from the autoPET-I winning weights and extended from two to four input channels, with the two scribble channels zero-initialised so that the pretrained representation is preserved exactly at initialisation. Every model is fine-tuned per fold from the corresponding autoPET-I fold checkpoint, so that no validation case is seen during pretraining. PET intensities are normalised against a per-scan aorta blood-pool reference derived from a CT segmentation, which removes tracer- and centre-specific scaling without requiring lesion labels. At inference the five fold models are ensembled by averaging their softmax outputs per sliding-window patch, before Gaussian-weighted stitching. On the challenge’s five-fold split, with each fold evaluated on its own validation cases, mean Dice is 0.554 and mean lesion-level F1 is 0.528 without scribbles, rising to 0.751 and 0.733 after five correction rounds. About 85% of that gain follows the first scribble, and the spread between fold models narrows five-fold over the same rounds, so interaction largely compensates for how well or badly a given model segments unaided. Keywords: autoPET challenge Interactive segmentation PET/CT 1 Introduction Positron emission tomography combined with computed tomography (PET/CT) is central to oncological staging and therapy response assessment, and quantitative read-out of tumour burden requires lesion segmentation. Manual delineation of whole-body studies is slow and subject to inter-reader variation, which motivates automated methods. The autoPET challenge series [11], built on a publicly released whole-body PET/CT cohort with manual lesion annotations [6, 7, 8], has driven progress on this task. Its fifth edition (autoPET/CT V) differs from the earlier editions in two respects. First, it is multitracer: the cohort contains both 18F-FDG and 68Ga/18F-PSMA studies, whose physiological uptake patterns differ substantially, so that a model tuned to one tracer does not transfer to the other without care. Second, and centrally, the task is interactive: for each study the challenge supplies user scribbles that mark lesion (foreground) and non-lesion (background) voxels, and the algorithm is expected to use them. The clinically relevant question is therefore not only how well a model segments unaided, but how reliably it responds when a reader points at what it got wrong. This paper describes our submitted algorithm. We treat the scribbles as two additional input channels to a three-dimensional residual encoder U-Net, initialise from the autoPET-I winning solution of Rokuss et al. [1], and fine-tune on the challenge data. We report five-fold cross-validation results over successive interactive correction rounds, which show that segmentation quality improves consistently as scribbles accumulate, most of the gain arriving with the first scribble. 2 Methods 2.1 Data Training data We use exclusively the training data provided by the challenge: 1611 whole-body PET/CT studies [6, 7, 8], comprising 1014 18F-FDG and 597 PSMA examinations; the PSMA cohort combines 369 studies acquired with an 18F-labelled and 228 with a 68Ga-labelled PSMA ligand; we use the v2 data release, which corrects a small number of segmentation masks relative to the initial release. Each study is accompanied by a manual lesion annotation. No external or private data is used. For each study the challenge additionally provides foreground and background scribbles; these are used as supplied and are described in Section 2.2. Validation data We follow the five-fold split distributed with the challenge data in splits_final.json, which divides the 1611 studies into folds of 1288/323 and 1289/322 training/validation cases. All results in Section 3 are obtained by evaluating each fold’s model on that same fold’s validation split, so that every one of the 1611 studies is predicted exactly once by a model that did not see it during training. The same split is used by the autoPET-I reference solution of Rokuss et al. [1], from which our weights are initialised, and we initialise each fold from that solution’s checkpoint for the same fold; consequently no validation case is seen during pretraining either. The per-fold checkpoints are those published by the authors via Zenodo [2]. 2.2 Data pre-processing All channels are resampled onto one common grid, defined from the original CT geometry: the voxel spacing is set to 2.0364×2.0364×3.02.0364× 2.0364× 3.0 m and the grid size to round(size⋅spacing/target spacing)round(size·spacing/target spacing) per axis, retaining the CT origin and direction. The identical procedure is applied at training time and inside the inference container. CT. Intensities are clipped to [−811.73,1137.42][-811.73,1137.42] HU and z-scored with the dataset-wide statistics μ=124.578μ=124.578, σ=280.883σ=280.883, then resampled with third-order B-spline interpolation and stored as float16. PET. Standardised uptake values are divided by a per-scan reference and passed through an inverse hyperbolic sine before a fixed dataset-wide z-score with μ=0.0904μ=0.0904, σ=0.2538σ=0.2538: PETnorm=(arcsinh(SUV/(r+ε))−μ)/σ,ε=10−8.PET_norm= (arcsinh(SUV/(r+ ))-μ )/σ, =10^-8. The reference r is the robust blood-pool activity of the aorta. An aortic mask is obtained by running TotalSegmentator [5] (task="total", fast=True) on the CT, transferring the mask to the PET grid by nearest-neighbour interpolation and eroding it by one voxel to suppress partial-volume contamination from the vessel wall. Within the mask, focal outliers are removed at median+3⋅1.4826⋅MADmedian+3· 1.4826·MAD and r is the mean of the remaining voxels. If the eroded mask contains fewer than 100 voxels with positive uptake, the reference falls back to the median of soft tissue (−100<HU<100-100<HU<100), and failing that to the median over all positive PET voxels. This makes the normalisation label-free and tracer-agnostic. The normalised volume is resampled with B-spline interpolation and stored as float16. Scribbles. The scribble heat maps are used exactly as provided by the challenge and are not regenerated. In the reference implementation they are produced by sampling up to k=5k=5 connected components of the annotation (26-connectivity). For each sampled component the single axial slice in which it occupies the most voxels is selected, and one scribble is drawn on that slice with the default centreline strategy, one of three strategies in the reference simulator, whose designs follow ScribblePrompt [9] and nnInteractive [10]: the two-dimensional slice mask is skeletonised, the largest skeleton component is retained (8-connectivity), the longest path between two skeleton endpoints is extracted and truncated by a fraction of 0.10.1. Background scribbles are drawn analogously inside a thin shell obtained by twice dilating the component with a unit-radius ball and subtracting the component itself – a two-voxel ring hugging the lesion boundary. At evaluation time background scribbles instead mark false-positive regions of the prediction, so the training and test semantics of the background channel differ; the interactive results in Section 3 measure the net effect of this mismatch. Each scribble is therefore a planar curve lying in one axial slice, not a three-dimensional structure. The resulting scribble voxels are written into a volume on the PET grid with value 11 and convolved with a Gaussian of σ=0σ=0, which is an identity operation; the maps are therefore binary point maps rather than smoothed heat maps, and are extremely sparse (of the order of 10−610^-6 of the volume by occupancy). Foreground and background scribbles are stored as two separate volumes. We resample them to the target spacing with linear interpolation, clip to [0,1][0,1], and quantise to uint8 as ⌊255x⌉ 255x ; they are rescaled to [0,1][0,1] by division by 255 when a patch is loaded. Annotations. Lesion masks are resampled to the target spacing by nearest-neighbour interpolation. 2.3 Algorithm/model The network is a three-dimensional residual encoder U-Net following the nnU-Net [4] ResEncL preset, configured with four input channels, two output classes and deep supervision enabled. The four channels are CT, normalised PET, the foreground scribble map and the background scribble map, in that order. Weights are initialised from the autoPET-I winning solution of Rokuss et al. [1], which is itself pretrained on a large multi-dataset corpus (Dataset619_nativemultistem) and then fine-tuned on the autoPET-I cohort. Two adaptations are made. First, the input stem is extended from the two channels of the original model (CT and PET) to four; the convolution weights of the two additional scribble channels are zero-initialised, so that at initialisation the extended network computes exactly the function of the pretrained two-channel model and the scribble pathway is learned from scratch during fine-tuning. Second, the pretrained model carries an auxiliary organ segmentation head, which we disable during fine-tuning: the organ information is assumed to be retained in the encoder representation and is not supervised further. Fold i of our model is initialised from fold i of the pretrained model, so the fold structure is preserved end to end. 2.4 Data post-processing Inference is performed by sliding a 192×192×192192× 192× 192 window over the resampled volume with an overlap of 0.50.5, i.e. a stride of 9696 voxels along each axis, with the last window on each axis shifted inward so that it fits inside the volume. Should an axis be shorter than the patch, the volume is zero-padded at its far end and the result cropped back afterwards. The forward passes run in mixed precision, as during training. The per-patch class posteriors, obtained by a softmax over the network logits, are accumulated into the output volume with a Gaussian importance weight, which down-weights the poorly contextualised patch borders relative to the patch centre. The weight map is obtained by placing a unit impulse at the patch centre, convolving it with a Gaussian whose standard deviation is σscale=0.125 _scale=0.125 of the patch size along each axis (i.e. 2424 voxels for a patch of 192192), normalising by its maximum and clipping from below at 10−310^-3. Two accumulators are maintained: the sum of weighted class probabilities and the sum of the weights themselves; their element-wise ratio yields the final probability volume, which is correct irrespective of how many windows cover a given voxel. The foreground channel of this probability volume is then resampled back to the original acquisition grid by linear interpolation and clipped to [0,1][0,1], and only afterwards thresholded at 0.650.65 to produce the binary lesion mask, which is written on the reference volume’s grid and affine. Applying the threshold after rather than before resampling avoids nearest-neighbour interpolation of a binary mask, which would otherwise discard small lesions. The threshold was selected by a sweep on fold 0’s validation split and applied unchanged to all folds and to the ensemble. No connected-component filtering, size thresholding or other morphological post-processing is applied. 2.5 Training and test parameters Each fold is trained for 500 epochs with stochastic gradient descent at a learning rate of 10−310^-3, Nesterov momentum of 0.990.99, zero dampening and a weight decay of 3×10−53× 10^-5. No learning rate schedule is used. The batch size is 22, training runs in mixed precision, and gradients are clipped to a maximum norm of 1212. The loss is the sum of cross-entropy and a soft Tversky term, ℒ=CE+(1−1|C|∑c∈CTPc+ϵTPc+αFPc+βFNc+ϵ),L=CE+ (1- 1|C| _c∈ C TP_c+εTP_c+α\,FP_c+β\,FN_c+ε ), with α=0.3α=0.3, β=0.7β=0.7 and ϵ=10−5ε=10^-5. The Tversky term is averaged over foreground classes only; including the background would cancel the asymmetry between α and β, since a foreground false negative is a background false positive. Penalising false negatives more heavily than false positives reflects the clinical cost of missing a lesion. The background remains supervised through the cross-entropy term. Deep supervision is applied with the standard nnU-Net weighting [1,12,14,…][1, 12, 14,…], normalised to unit sum and with the coarsest resolution level dropped. Training patches are sampled stochastically rather than augmented. For a tumour-bearing case, with probability 0.330.33 a hard-negative, tumour-free window is drawn (the first such window found within 20 attempts, otherwise the window containing the fewest lesion voxels); otherwise a lesion is drawn uniformly at the lesion level, independently of its size, and a window is placed at random such that it contains a random voxel of that lesion. Tumour-free cases receive a uniformly random window. No intensity or geometric augmentation of any kind is applied. For every fold the checkpoint with the lowest validation loss over the 500 epochs is selected, using the same criterion for all five folds. Training was performed on a single NVIDIA V100 with 32 GB of memory. At test time no test-time augmentation is used. The five fold models are ensembled by averaging their softmax outputs per sliding-window patch, before the patch is stitched into the output volume, so that the Gaussian weighting and reverse resampling described above operate on a single already-ensembled probability map. 2.6 Github repository Link to Github repository: https://github.com/amged-elmasry-bs/bs_autopetV_submission 3 Results We report five-fold cross-validation on the challenge training data using the splits_final.json split, evaluating each fold’s model on that fold’s validation cases. All scores are produced by the challenge’s official MetricEvaluator [3], using its two metrics unmodified. The first is the volumetric Dice between the predicted mask P and the annotation G, Dice=2|P∩G||P|+|G|.Dice= 2\,|P∩ G||P|+|G|. It is evaluated on the original acquisition grid. When the annotation contains no lesion, |G|=0|G|=0, the quotient is undefined and the evaluator returns NaN. This depends on the annotation alone: a lesion-free study yields NaN whether the model correctly predicts nothing or produces a false positive, so such cases carry no Dice at all rather than a score of zero or one. The second is a lesion-level F1F_1. Both volumes are decomposed into connected components at 18-connectivity, and a predicted component is counted as a true positive when its intersection over union with an annotated lesion is at least 0.10.1; unmatched predicted components are false positives and unmatched annotated lesions are false negatives. With those lesion counts, F1=2TP2TP+FP+FN.F_1= 2\,TP2\,TP+FP+FN. Every lesion contributes equally regardless of its size, and the 0.10.1 overlap requirement means the metric asks whether a lesion was detected rather than how precisely it was delineated. F1F_1 is likewise undefined when the annotation contains no lesion, since TP+FN=0TP+FN=0. Studies without annotated lesions are therefore excluded from both averages, leaving 1040 of the 1611 studies scored, between 199 and 215 per fold. Two further conventions matter for comparability. First, alongside the per-case macro average reported here, the evaluator pools lesion counts across all cases into an aggregated F1F_1, in which false positives on tumour-free studies do count. Second, the challenge ranks by the area under each metric’s six-round curve, a trapezoid over the round index in which the endpoint rounds r0r_0 and r5r_5 carry half the weight of the interior ones. We report it normalised by the number of intervals, so that it lies on the same [0,1][0,1] scale as the metric it integrates; per-fold AUCs are given in Tables 1 and 2. The reported figures follow the same convention as the official challenge leaderboard, which reports a count of scored cases smaller than the number submitted for exactly this reason. Note that these are single-model results: each fold model is scored only on the cases it did not train on. The five-fold ensemble that constitutes the actual submission cannot be evaluated in this protocol, since for any training case four of its five members have seen that case; its performance is measured on the held-out challenge test set. Performance is reported over successive interactive rounds. Round r0r_0 is the unaided prediction, with both scribble channels empty. In each subsequent round one scribble is generated from the model’s current errors and appended to the accumulated input, following the challenge’s reference interactive loop: a candidate scribble is drawn on the false-positive region (y^=1 y=1, y=0y=0) and another on the false-negative region (y^=0 y=0, y=1y=1), both by the challenge’s own scribble generator. For each error region it examines every axial slice, takes the largest two-dimensional connected component (8-connectivity) within each, and selects the slice whose component is largest overall; the scribble is the centreline of that component and lies entirely in that one slice. Of the two candidates the longer is kept – a foreground scribble when the false-negative scribble is at least as long as the false-positive one, and a background scribble otherwise. A round therefore adds a single planar curve, typically a few voxels in extent, to whichever error the generator could trace furthest. Scribbles are encoded exactly as at training time and as in the submitted container: written as ones on the original grid, resampled linearly to the target grid, clipped and quantised through uint8, with no additional smoothing or renormalisation. Each case therefore receives six inference passes in total: r0r_0 without any scribble, followed by five refinement rounds r1…r5r_1… r_5, each adding one further scribble to those already accumulated. Table 1: Five-fold cross-validation, Dice over the six interactive rounds (r0r_0 without scribbles, then five refinement rounds). Each fold’s model is evaluated on its own validation split. Both the per-round means and the AUC are computed over tumour-bearing cases only, since Dice is undefined on a lesion-free annotation and the evaluator returns NaN. The AUC is the trapezoid of the six-round curve over the round index, normalised by the number of intervals so that it shares the scale of the metric itself and its maximum is 1.01.0. The Mean row averages the five fold curves. Fold r0r_0 r1r_1 r2r_2 r3r_3 r4r_4 r5r_5 AUC 0 0.5708 0.7344 0.7536 0.7591 0.7637 0.7654 0.7358 1 0.4583 0.7130 0.7424 0.7523 0.7598 0.7638 0.7157 2 0.5374 0.7164 0.7242 0.7281 0.7300 0.7310 0.7066 3 0.6312 0.7291 0.7422 0.7480 0.7528 0.7540 0.7330 4 0.5719 0.7186 0.7328 0.7390 0.7414 0.7416 0.7177 Mean 0.5539 0.7223 0.7391 0.7453 0.7495 0.7512 0.7218 Table 2: Five-fold cross-validation, lesion-level F1F_1 over the six interactive rounds. As in Table 1, the means and the AUC are computed over tumour-bearing cases only, F1F_1 being undefined when the annotation contains no lesion, and the AUC is the normalised trapezoid over the round index (maximum 1.01.0). Fold r0r_0 r1r_1 r2r_2 r3r_3 r4r_4 r5r_5 AUC 0 0.5327 0.7160 0.7359 0.7388 0.7493 0.7506 0.7163 1 0.4394 0.6908 0.7139 0.7232 0.7302 0.7395 0.6895 2 0.5103 0.7010 0.7066 0.7093 0.7122 0.7135 0.6882 3 0.6142 0.7062 0.7250 0.7272 0.7334 0.7347 0.7133 4 0.5437 0.7039 0.7178 0.7230 0.7224 0.7246 0.7002 Mean 0.5281 0.7036 0.7198 0.7243 0.7295 0.7326 0.7015 Averaged over the five folds, the unaided model reaches a Dice of 0.55390.5539 and a lesion-level F1F_1 of 0.52810.5281. A single scribble raises these to 0.72230.7223 and 0.70360.7036, and after five rounds they stand at 0.75120.7512 and 0.73260.7326 – a total gain of +0.1973+0.1973 Dice and +0.2045+0.2045 F1F_1. The gain is heavily front-loaded: the first scribble alone accounts for 85%85\% of it in both metrics, with the remaining four rounds contributing +0.0289+0.0289 Dice and +0.0290+0.0290 F1F_1 between them. The curves are therefore steep at r1r_1 and close to flat by r3r_3, which is what the normalised AUC of 0.72180.7218 (Dice) and 0.70150.7015 (F1F_1) reflects. Dice increases at every round on all five folds. F1F_1 does so on four of them; fold 4 dips by 0.00060.0006 between r3r_3 and r4r_4 before recovering, an excursion far smaller than the between-fold spread and the only non-monotonic step in the sixty transitions recorded. At the level of individual studies, 497497 of the 10401040 scored cases improve in F1F_1 from r0r_0 to r5r_5, 461461 are unchanged and 8282 deteriorate. The most pronounced effect is on the spread between folds. Unaided, the folds differ by 0.17290.1729 Dice (0.45830.4583 on fold 1 to 0.63120.6312 on fold 3); after five rounds that spread is 0.03440.0344 (0.73100.7310 to 0.76540.7654), a five-fold compression, with the corresponding F1F_1 spread falling from 0.17480.1748 to 0.03710.0371. The ordering is not preserved either: fold 1 is the weakest fold unaided and the second strongest at r5r_5, whereas fold 2 – whose selected checkpoint also has the highest validation loss of the five – gains least after the first scribble and ends last. Interaction thus compensates for much of the variation in unaided quality between fold models, which is the behaviour one wants from a corrective interface. 4 Discussion The cross-validation results in Tables 1 and 2 show that segmentation quality increases with each interactive round: the model follows the supplied scribbles and progressively corrects its own errors rather than ignoring the additional channels or degrading as more input is accumulated. Averaged over folds the loop is worth +0.1973+0.1973 Dice and +0.2045+0.2045 F1F_1, and the first scribble supplies roughly 85%85\% of that; the four subsequent rounds refine rather than transform the segmentation. Where the gain comes from is worth separating out. Pooled over all folds and cases, five rounds convert 829829 false negatives into true positives, while the false positive count rises slightly, from 68696869 to 69986998. The improvement is thus almost entirely a matter of recovering lesions the model had missed, not of suppressing spurious ones. This is despite the loop spending most of its budget on the latter: under the challenge’s rule the longer of the two candidate scribbles is kept, and because false-positive regions are typically larger and more elongated than the residual false-negative ones, 69%69\% of the 54265426 scribbles drawn are background scribbles. Those background scribbles are, on this evidence, close to inert. The likely reason is the semantic mismatch noted in Section 2.2: during training the background channel carries a thin shell hugging a true lesion, marking its boundary, whereas at evaluation it marks a whole false-positive region, and the network has never been trained to read it that way. Training with evaluation-style background scribbles – drawn on the model’s own false positives rather than around annotated lesions – is the most promising single change suggested by these results. Two design choices are worth noting. Normalising PET against a per-scan aorta blood-pool reference is label-free and does not depend on the tracer, which matters for a cohort combining FDG and PSMA studies whose absolute uptake scales differ. Zero-initialising the two scribble channels means the extended four-channel network reproduces the pretrained two-channel model exactly before any fine-tuning, so the interactive capability is added without disturbing the segmentation behaviour inherited from the autoPET-I solution. Several limitations should be stated. The cross-validation figures are single-model results; the submitted five-fold ensemble cannot be scored under this protocol, so its benefit is assumed from the usual behaviour of nnU-Net ensembles rather than measured here. The threshold of 0.650.65 was chosen on fold 0 and reused unchanged everywhere, including for the ensemble, whose probability distribution is not identical to a single model’s. Scores are reported on tumour-bearing studies only, because both metrics are undefined without an annotated lesion; this excludes 571571 of the 16111611 studies and means that false positives on lesion-free scans, clinically the more costly error, do not enter the per-case averages at all. The simulated reader is also an optimistic one: scribbles are generated automatically from the ground truth by the challenge’s own simulator, always mark a genuine error, and are drawn as clean centrelines, so the results bound what a perfectly informed annotator would obtain rather than what a human would. Finally, the background channel behaves as described above, which caps what further rounds can achieve. 5 Conclusion We described our submission to the autoPET/CT V challenge: a scribble-conditioned three-dimensional residual encoder U-Net, initialised per-fold from the autoPET-I winning weights, extended from two to four input channels with zero-initialised scribble pathways, trained with a false-negative-weighted Tversky and cross-entropy objective, and deployed as a five-fold ensemble whose fold predictions are averaged per sliding-window patch. Five-fold cross-validation shows that performance improves consistently across successive interactive rounds, confirming that the model uses the supplied scribbles to correct its own errors. Averaged over the five folds, Dice rises from 0.55390.5539 unaided to 0.75120.7512 after five scribbles and lesion-level F1F_1 from 0.52810.5281 to 0.73260.7326, with about 85%85\% of the gain delivered by the first scribble, and the spread between fold models narrows five-fold over the same rounds. Table 3: Algorithm details Team name algorithm name data pre-processing data post-processing training data augmentation BS BS baseline Resampled to a common grid derived from the CT geometry at 2.0364 x 2.0364 x 3.0 m. CT: clip to [−811.7,1137.4][-811.7,1137.4] HU, dataset z-score, B-spline. PET: arcsinharcsinh(SUV / per-scan aorta blood-pool reference), dataset z-score, B-spline; reference from a TotalSegmentator aorta mask with MAD outlier rejection. Scribbles: linear resample, clip [0,1][0,1], uint8. Sliding window 1923192^3 at 0.50.5 overlap; per-patch softmax stitched with Gaussian importance weights (σ=0.125×σ=0.125× patch); reverse linear resample to the original grid; threshold 0.650.65. No connected-component or size filtering. None. Stochastic patch sampling only (pbg=0.33p_bg=0.33 hard negatives, otherwise per-lesion uniform) test time augmentation ensembling standardized framework network architecture loss None 5-fold cross-validation ensemble; softmax averaged per sliding-window patch, before Gaussian stitching nnU-Net, ResEncL preset (3D) Residual Encoder U-Net (3D); 4 input channels (CT, PET, foreground and background scribbles), 2 output classes, deep supervision Cross-entropy ++ Tversky (α=0.3α=0.3, β=0.7β=0.7, foreground classes only), deep-supervision weighted training data data/model dimensionality and size use of pre-trained models GPU hardware for training 1014 FDG ++ 597 PSMA PET/CT of autoPET (369 18F- and 228 68Ga-labelled PSMA); 5-fold split of splits_final.json 3D; patch 192×192×192192× 192× 192 at 2.0364 x 2.0364 x 3.0 m Public: autoPET-I winning model of Rokuss et al. [1, 2], itself pretrained on Dataset619_nativemultistem. Fold i initialised from their fold i; input stem extended 2 → 4 channels with the two scribble channels zero-initialised; auxiliary organ head disabled 1x NVIDIA V100 (32 GB) Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. All authors are employees of Brightskies. No author is affiliated with any of the institutes of the challenge organizers. References [1] Rokuss, M., Kovacs, B., Kirchhoff, Y., Xiao, S., Ulrich, C., Maier-Hein, K.H., Isensee, F.: From FDG to PSMA: A Hitchhiker’s Guide to Multitracer, Multicenter Lesion Segmentation in PET/CT Imaging. arXiv:2409.09478 (2024) [2] Rokuss, M.: Model checkpoint of the autoPET I LesionTracer solution. Zenodo (2024). https://doi.org/10.5281/zenodo.14007247. Code: https://github.com/MIC-DKFZ/autopet-3-submission. Last accessed 26 Aug 2026 [3] autoPET/CT V challenge: official evaluation code and interactive baseline. https://github.com/lab-midas/autoPETV. Last accessed 26 Aug 2026 [4] Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203–211 (2021) [5] Wasserthal, J., Breit, H.-C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., Bach, M., Segeroth, M.: TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology: Artificial Intelligence 5(5), e230024 (2023) [6] Gatidis, S., Hepp, T., Früh, M., La Fougère, C., Nikolaou, K., Pfannenberg, C., Schölkopf, B., Küstner, T., Cyran, C., Rubin, D.: A whole-body FDG-PET/CT dataset with manually annotated tumor lesions. Scientific Data 9, 601 (2022). https://doi.org/10.1038/s41597-022-01718-3 [7] Gatidis, S., Kuestner, T.: A whole-body FDG-PET/CT dataset with manually annotated tumor lesions (FDG-PET-CT-Lesions) [dataset]. The Cancer Imaging Archive (2022). https://doi.org/10.7937/gkr0-xv29 [8] Jeblick, K., Schachtner, B., Mittermeier, A., Dexl, J., Wesp, P., Küstner, T., Gatidis, S., Früh, M., Fabritius, M., Unterrainer, L., Sheikh, G., Delker, A., Böning, G., Brendel, M., Ricke, J., Werner, R., Gu, S., Ingrisch, M., Geyer, T., Cyran, C.: A whole-body PSMA-PET/CT dataset with manually annotated tumor lesions (PSMA-PET-CT-Lesions) (Version 1) [dataset]. The Cancer Imaging Archive (2024). https://doi.org/10.7937/r7ep-3x37 [9] Wong, H.E., Rakic, M., Guttag, J., Dalca, A.V.: ScribblePrompt: Fast and Flexible Interactive Segmentation for Any Biomedical Image. In: ECCV 2024. arXiv:2312.07381 [10] Isensee, F., Rokuss, M., Krämer, L., et al.: nnInteractive: Redefining 3D Promptable Segmentation. arXiv:2503.08373 (2025) [11] Kuestner, T., Gatidis, S., Megne, O., Peisen, F., Ingrisch, M., Fabritius, M., Dexl, J., Jeblick, K., Cyran, C., Shiyam Sundar, L.K., Marinov, Z., Jackson, P., Hofman, M., Kleesiek, J., Kim, M., Herrmann, K.: AutoPET: Automated Lesion Segmentation in Whole-Body PET/CT (v1). Zenodo (2026). https://doi.org/10.5281/zenodo.19714420