Source-linked AI summary
Three-Phase Scribble-Adaptive Curriculum Learning for autoPETV Grand Challenge
Libo Zhang
TL;DR
Interactive whole-body PET/CT lesion segmentation requires models to refine initial masks from sparse corrective scribbles. The report introduces a three-phase curriculum that encodes scribbles in two channels and trains a residual-encoder U-Net with simulated guidance and self-error correction. In five-fold interactive evaluation, the final checkpoints improve monotonically, with most Dice improvement arriving after the first correction.
Problem
autoPETV benchmarks interactive lesion segmentation in whole-body PET/CT, where an initial mask must be refined through sparse corrective scribbles over multiple iterations.
Method
A residual-encoder U-Net receives accumulated foreground and background scribbles through two extra channels and is trained with a three-phase curriculum culminating in online error-driven multi-step simulation.
Results
The final checkpoints improve monotonically in every fold; the first corrective scribble contributes about 55 percent of total Dice gain, from 0.657 to 0.742.
Takeaways & Limitations
Minimal guidance produces most of the improvement, while the learned corrective behavior remains consistent across folds, tracers, and simulated scribble styles.
Takeaways & Limitations
Binary unsmoothed interaction heatmaps and threshold-based DeepPSMA labels are deliberate simplifications, with the latter treated as tolerable training noise.
Abstract
from arXiv · showhide
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.
1 Introduction
The autoPETV challenge evaluates interactive whole-body PET/CT lesion segmentation, where an initial mask is refined through sparse corrective scribbles. This submission represents those scribbles as additional channels and trains with a curriculum aligned to iterative correction.
- Challenge setting: Interactive autoPETV segmentation refines an initial lesion mask as sparse corrective scribbles arrive over several iterations.Performance uses the area under per-step Dice and detection-matching-metric curves.
- Submission approach: The method encodes accumulated foreground and background scribbles in two extra input channels and trains through online, error-driven multi-step simulation.DeepPSMA data is added to every fold’s training split, and best and final checkpoints from five folds are ensembled at test time.
2 Methods
The method combines autoPET and DeepPSMA training data with two rasterized scribble channels, a three-phase curriculum, and a ten-checkpoint ensemble for interactive inference. Training simulates both ground-truth-derived guidance and corrections targeted at the model’s own errors.
- Data: 1811 PET/CT studies are used for training: 1014 autoPET FDG, 597 autoPET PSMA, and 200 DeepPSMA studies.All DeepPSMA studies enter every fold’s training split, while validation remains the official reference split.
- Data: Five official validation folds contain 322 or 323 original autoPET studies each and exclude DeepPSMA data.This preserves the official validation reference split.
- Interaction representation: Each case uses CT, PET SUV, and two binary channels encoding accumulated foreground and background scribbles.The interaction history enters the segmentation network through heatmap channels, while predictions are stateless at each round.
- Network: The residual-encoder U-Net uses an extra-large nnU-Net configuration with seven stages and about 140 million parameters.The planned compute budget is 40 GB, with 192×256×256 patches and deep supervision.
- Curriculum: The three-phase curriculum spans 4000 epochs, progressing from silent interaction channels to masked ground-truth scribbles and error-driven self-adaptation.Phase A covers epochs 0–200, Phase B epochs 200–3200, and Phase C the remaining epochs.
- Curriculum: Phase C samples up to five correction steps on the currently larger error class, then rasterizes their scribbles before the gradient-carrying pass.Two forward passes per sample are used irrespective of correction depth.
- Inference: The submission averages logits from the best and final checkpoints of all five folds before taking the arg-max.This produces a ten-checkpoint ensemble for prediction.
- Optimization: Training uses SGD for 4000 epochs with Nesterov momentum 0.99, weight decay 3×10−5, and a polynomial learning-rate schedule.The initial learning rate is 5×10−4 and the schedule exponent is 0.9.
3 Results
Interactive evaluation used six steps per case and mirrored the training simulation, with performance measured by trapezoidal AUC-Dice and AUC-DMM across five folds.
- Evaluation protocol: Six-step evaluation applied one corrective scribble per step to the currently larger error class.Each case received an automatic prediction followed by five corrections.
- Evaluation protocol: Scribble strategies were assigned round-robin so each strategy covered one third of every validation fold.Lesion-level detection matched 18-connected components at an intersection-over-union threshold of 0.1.
- Evaluation protocol: AUC-Dice and AUC-DMM were computed as trapezoidal areas over per-case metric values from interaction steps 0 through 5.The per-case metric M was either Dice or DMM.
- Results: All ten runs improved monotonically, while the first corrective scribble delivered about 55 percent of the total Dice gain.For final checkpoints, five-fold mean Dice increased from 0.657 initially to 0.742 after one step and 0.810 after five.
- Results: The final checkpoint was better in nine of ten fold-metric comparisons, but the best checkpoint produced the strongest single run.The strongest run was fold 4 with AUC-Dice 3.898.
4 Discussion
The discussion attributes front-loaded interaction gains to error-driven Phase C training and reports complementary tracer behavior and checkpoint performance that motivate ensembling.
- Interaction gains: Minimal guidance produced most of the improvement, with the first correction supplying the largest interaction gain.The discussion links this front-loaded behavior to Phase C training on the network’s own largest residual errors.
- Tracer profiles: The learned correction mechanism addressed volumetric refinement of FDG lesions and detection of numerous small PSMA lesions without tracer-specific tuning.This interpretation is presented as a consequence of the complementary tracer profiles.
- Checkpoint complementarity: The expected benefit of adding the extra checkpoints over a five-checkpoint ensemble was small but consistently non-negative.Logit averaging did not alter the interaction protocol or per-invocation time-limit compliance.
5 Conclusion
A three-phase curriculum produces a scribble-adaptive residual-encoder nnU-Net that improves monotonically under corrective guidance across folds, tracers, and simulated scribble styles.
- Conclusion: A three-phase curriculum turns a standard residual-encoder nnU-Net into a scribble-adaptive segmenter.The model improves monotonically under corrective guidance.
- Conclusion: Two additional input channels provide interaction without architectural modification beyond the scribble representation.The underlying self-configuring pipeline therefore carries over to the interactive setting.
- Conclusion: Observed gains were consistent across folds, tracers, and simulated scribble styles.The report presents this consistency as indicating that corrective behavior is tied to the training regime rather than a particular annotation pattern.
A Training Curves
Figures 2–6 show training progress across all five folds. Phase C begins at epoch 3200, adding a gradient-free forward pass that roughly doubles each epoch’s duration.
- Phase C: Epoch 3200 marks the onset of Phase C across the five-fold training progress.Phase C introduces an additional gradient-free forward pass.
- Phase C: Phase C roughly doubles the duration of every training epoch.The added forward pass is responsible for the increased epoch duration.