Three-Phase Scribble-Adaptive Curriculum Learning for autoPETV Grand Challenge

Libo Zhang; Yue Ning

cs.CVAug 22, 2026

Three-Phase Scribble-Adaptive Curriculum Learning for autoPETV Grand Challenge

Authors: Libo Zhang, Yue Ning

Organizations: Duke University

Abstract

This report describes Libo Zhang's algorithmic solution to autoPETV Grand Challenge on interactive lesion segmentation in whole-body PET/CT. Interaction is encoded as two additional input channels that rasterize the accumulated foreground and background scribbles, and a residual-encoder U-Net of about 140 million parameters is trained with a three-phase curriculum over 4000 epochs: the network first learns fully automatic segmentation with silent interaction channels, then observes ground-truth-derived scribbles under randomly sampled visibility modes, and finally adapts to its own mistakes through online simulation of up to five error-driven correction steps. Training draws on 1811 autoPET and DeepPSMA studies, and the submission ensembles the best and final checkpoints of five folds by logit averaging. In interactive five-fold cross-validation with six interaction steps, the final checkpoints reach a mean AUC-Dice of 3.836 and a mean AUC-DMM of 3.869, improving monotonically in every fold, with roughly half of the total gain delivered by the first corrective scribble. Our code and trained model checkpoints are available on https://github.com/Libo1023/autoPETV-Curriculum.

Figures & tables

Team name	LiboZhang
Algorithm name	autoPETV-Curriculum
Data pre-processing	DeepPSMA CT linearly resampled to the PET grid; all channels resampled to the median spacing of $3.0\times 2.04\times 2.04$ mm (spline order 3 for images, linear for labels); CT clipped to $[-822.5,1127.0]$ and z-scored (global mean 126.7, standard deviation 286.7); PET and both interaction channels z-scored per image; scribbles rasterized as two binary heatmaps
Data post-processing	Arg-max of the ensemble-averaged logits
Training data augmentation	Default nnU-Net pipeline on all four channels (rotation $\pm 30^{\circ}$ , scaling 0.7 to 1.4, Gaussian noise and blur, brightness and contrast 0.75 to 1.25, simulated low resolution, gamma, mirroring on all axes); three-phase interaction curriculum of Sect. 2.3
Standardized framework	nnU-Net v2 (2.6.0) on PyTorch 2.6.0 with CUDA 12.4
Network architecture	3D residual-encoder U-Net at a 40 GB compute budget: 7 stages with 32 to 320 features, encoder depths $(1,3,4,6,6,6,6)$ , deep supervision, about 140M parameters, each checkpoint’s size is roughly 1.1 GB

Table 1: Algorithm details.

Figure 1: Per-step mean Dice (left column) and DMM (right column) over the six-step interactive evaluation across five folds. Each panel overlays the best (dashed) and final (solid) checkpoints for all lesion-positive cases (circles) and for the FDG (squares) and PSMA (triangles) subsets; the AUC-Dice and AUC-DMM of both checkpoints are annotated per panel. Metrics are the areas under these curves following Eq. ( 4 ).

		AUC-Dice			AUC-DMM
Fold ( $n$ )	Ckpt	All	FDG	PSMA	All	FDG	PSMA
0 (207)	B	3.8739	4.0525	3.7070	3.9011	3.6764	4.1110
	F	3.9087	4.0936	3.7359	3.9581	3.7940	4.1114
1 (215)	B	3.7320	4.0672	3.4565	3.8039	3.6299	3.9470
	F	3.7532	4.0726	3.4907	3.8490	3.7181	3.9565
2 (213)	B	3.7793	4.1725	3.4111	3.7941	3.7117	3.8712

Table 2: Interactive five-fold cross-validation of the best (B) and final (F) checkpoints, overall and per tracer.

n

is the number of lesion-positive validation cases of each fold, and the better checkpoint per fold and column is highlighted in bold.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net
Sep 1, 2026Marven Sherif, Amgad Elmasry, Youssef Ghazal +1Lesion Segmentation Positron Emission Tomography
Brightskies
Pretrained, Curriculum-Tuned, and Ensembled: A Tracer-Aware Interactive Segmentation Pipeline for AutoPET V
Aug 31, 2026Xinglong Liang, Chunyao Lu, Tianyu Zhang +4Lesion Segmentation Multi Organ Computed Tomography Segmentation
Department of Radiology, The Netherlands Cancer Institute, Plesmanlaan 121, 1066 CX, Amsterdam, the Netherlands · Department of Radiology and Nuclear Medicine, Radboud University Medical Centre, Geert Grooteplein 10, 6525 GA, Nijmegen, The Netherlands · Faculty of Applied Sciences, Macao Polytechnic University, 999078, Macao Special Administrative Region of China +2
The autoPET3 Challenge: Automated Lesion Segmentation in Whole-Body PET/CT $\unicode{x2013}$ Multitracer Multicenter Generalization
May 7, 2026Jakob Dexl, Katharina Jeblick, Andreas Mittermeier +27Lesion Segmentation Positron Emission Tomography
Department of Radiology, LMU University Hospital, LMU Munich, Munich, Germany · Munich Center for Machine Learning (MCML), Munich, Germany · Comprehensive Pneumology Center (CPC-M), Member of the German Center for Lung Research (DZL), Munich, Germany +13

Sep 1, 2026, cs.CVJ/K move · Enter open · S save

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

Brightskies

Lesion Segmentation Positron Emission Tomography U-Net Cone-Beam Computed Tomography Intermediate Checkpoints

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net

Pretrained, Curriculum-Tuned, and Ensembled: A Tracer-Aware Interactive Segmentation Pipeline for AutoPET V

The autoPET3 Challenge: Automated Lesion Segmentation in Whole-Body PET/CT \unicodex2013\unicode{x2013}\unicodex2013 Multitracer Multicenter Generalization

The autoPET3 Challenge: Automated Lesion Segmentation in Whole-Body PET/CT $\unicode{x2013}$ Multitracer Multicenter Generalization