Source-linked AI summary
Pretrained, Curriculum-Tuned, and Ensembled: A Tracer-Aware Interactive Segmentation Pipeline for AutoPET V
Xinglong Liang, Chunyao Lu, Tianyu Zhang, Jiaju Huang, Tao Tan, Yunchao Yin, Lishan Cai
TL;DR
Whole-body PET/CT interactive segmentation must produce reliable initial masks and incorporate sparse corrections despite differing FDG and PSMA uptake patterns. TRIAGE combines masked-autoencoder-pretrained STU-Net features, anatomical organ context, tracer-specific branches, and scribble-conditioned second-stage refinement. Ten-fold cross-validation reported higher mean Dice after refinement for both tracers, increasing from 0.5889 to 0.6175 for FDG and from 0.5833 to 0.6405 for PSMA.
Problem
Interactive whole-body PET/CT segmentation requires reliable initial predictions and efficient refinement from sparse scribbles across substantially different FDG and PSMA imaging characteristics.
Method
TRIAGE uses MAE-pretrained STU-Net, organ segmentation context, tracer-specific FDG and PSMA branches, and a scribble-conditioned second-stage refinement network.
Results
0.0286 and 0.0572 absolute Dice improvements were reported for FDG and PSMA, respectively, after second-stage refinement in ten-fold cross-validation.
Takeaways & Limitations
Second-stage anatomical and interaction guidance improved lesion segmentation over initial automatic predictions for both tracer domains, with a larger improvement for PSMA.
Abstract
from arXiv · showhide
Interactive lesion segmentation in whole-body PET/CT requires a model to provide a strong initial prediction while also responding efficiently to sparse corrective scribbles during inference. This setting is particularly challenging because tracer distributions, physiological uptake patterns, lesion appearance, and acquisition characteristics differ substantially between FDG and PSMA studies. We present TRIAGE, Tracer-aware Refinement via Interactive Anatomy-Guided sEgmentation. The core backbone is a 3D STU-Net initialized through masked autoencoding pre-training with an asynchronous masking strategy, aiming to learn transferable anatomical and cross-modal representations before task-specific fine-tuning. In parallel, we train an auxiliary organ segmentation model whose predictions provide explicit anatomical context and help distinguish physiological uptake from malignant lesions. A dedicated tracer classifier first routes each study to an FDG- or PSMA-specific branch. Within each branch, a first-stage segmentation model consumes CT, PET, and organ context to generate an initial lesion mask. The initial prediction is then combined with cumulative foreground/background scribbles and refined by a second interactive segmentation network. The FDG and PSMA branches share the same overall processing pipeline but are trained independently to account for tracer-specific appearance and error modes. We additionally employ curriculum-style training and model ensembling to improve robustness across interaction steps and heterogeneous cohorts. Experiments are conducted using the official AutoPET V data and ten-fold split; quantitative results, ablations, and final test-set performance are left as placeholders to be completed after the challenge evaluation. Code: https://github.com/Liiiii2101/AUTOPET2026-MEDAI.
1 Introduction
TRIAGE addresses the difficulty of interactive whole-body PET/CT lesion segmentation, where models must produce reliable initial masks and refine sparse corrections across distinct FDG and PSMA domains. It combines pre-training, anatomical context, tracer-specific models, and interaction-aware refinement.
- Whole-body PET/CT segmentation is important for tumor-burden assessment and treatment-response evaluation but is difficult because lesions are heterogeneous, scattered, and sometimes small.
- Physiological tracer uptake can resemble pathological uptake, motivating anatomical priors and methods that combine PET with CT context.
- Interactive AutoPET V segmentation requires a reliable initial lesion mask, interpretation of positive and negative scribbles, and progressive improvement across interactions.
- The pipeline separately processes FDG and PSMA with the same overall architecture, then uses a second-stage model conditioned on images, the initial prediction, and cumulative scribble guidance.
- TRIAGE uses MAE-pretrained STU-Net features together with organ-level anatomical context to strengthen initial lesion segmentation and suppress physiological false positives.
- Curriculum-style training and model ensembling are used to improve robustness across interaction steps and heterogeneous cohorts.
2 Methods
The method pre-trains a STU-Net-Small backbone on paired PET/CT, routes cases to independently trained tracer-specific models, and performs two-stage lesion segmentation with anatomical and scribble guidance. It uses the official AutoPET V cohorts and tracer-specific preprocessing and post-processing.
- STU-Net-Small is pre-trained with masked autoencoding on paired PET and CT volumes using independently sampled spatial masks.PET and CT use different mask locations, with a masking ratio of 0.5 for each modality.
- A tracer classification network routes each case to an independently trained FDG or PSMA segmentation model sharing the same architecture and pipeline.
- The first stage combines PET, CT, and anatomical organ masks to generate an initial lesion prediction, while the second stage adds the initial mask and user scribbles for refinement.
- Curriculum-based refinement training and model ensembling are used during inference to improve robustness.
- The study uses only the official AutoPET V training dataset, containing 1,014 FDG studies from 900 patients and 597 PSMA studies from 378 patients.
- Tracer-specific preprocessing resamples FDG images to (3.0, 2.03, 2.03) mm with 128×128×128 patches and PSMA images to (4.07, 3.27, 4.07) mm with 112×192×112 patches.
- A post-processing step removes predicted lesion voxels below SUV thresholds of 1.5 g/mL for FDG and 1.0 g/mL for PSMA.
3 Results
Ten-fold cross-validation shows that interactive second-stage refinement improves Dice over the initial automatic segmentation for both FDG and PSMA, with a larger gain for PSMA. The reported fold statistics include an incomplete FDG training fold excluded from aggregate results.
- Ten-fold cross-validation separately evaluates validation Dice for first-stage automatic segmentation and second-stage refinement in FDG and PSMA cohorts.
- 0.5889 ± 0.0537 FDG mean validation Dice and 0.5833 ± 0.0293 PSMA mean validation Dice were obtained by the first-stage model.
- 0.6175 ± 0.0433 FDG mean Dice and 0.6405 ± 0.0270 PSMA mean Dice were obtained by the second-stage model.
- 0.0286 and 0.0572 absolute Dice improvements were reported for FDG and PSMA, respectively, after second-stage refinement.
- FDG fold 4 training was incomplete, so aggregate FDG statistics used the remaining nine folds.
- The second-stage strategy improves segmentation accuracy over the initial prediction for both tracer domains, with the larger improvement observed for PSMA.
4 Discussion
The two-stage strategy improves lesion segmentation across FDG and PSMA by adding anatomical and interaction cues, while tracer-specific models accommodate distinct uptake and error patterns. Pre-training also improves STU-Net performance without increasing model size or inference complexity.
- Refinement benefits: The second-stage model improves mean Dice from 0.5889 to 0.6175 for FDG and from 0.5833 to 0.6405 for PSMA.The refinement stage uses anatomical organ predictions, the initial segmentation, and foreground/background scribbles.
- Tracer-specific effects: PSMA shows the larger Stage-2 improvement, suggesting anatomical and interaction guidance may be particularly beneficial for this tracer.Independent FDG and PSMA models address tracer-specific uptake distributions, lesion appearances, and false-positive patterns.
- Pre-training and efficiency: Pre-training raises STU-Net validation Dice from 0.3890 to 0.4889, an absolute gain of 0.0999 over the same architecture trained from scratch.The ablation used a single FDG validation fold without organ context and was intended to compare architecture and pre-training under matched conditions.
- Pre-training and efficiency: STU-Net trained from scratch achieves comparable validation Dice to nnU-Net while using 14.55M parameters and 138.70 GFLOPs versus 30.79M and 526.29 GFLOPs.The pre-training gain was achieved without increasing model size or inference complexity.
- Overall interpretation: The results support combining tracer-specific specialization, anatomical context, and interaction-aware refinement for whole-body PET/CT lesion segmentation.The refinement-stage advantage is reported consistently across both tracer domains.
5 Conclusion
The paper presents a tracer-aware two-stage framework combining masked-autoencoder pre-training, tracer-specific models, anatomical guidance, and scribble-conditioned refinement. Ten-fold cross-validation reports improved mean Dice after refinement for both FDG and PSMA, while official test-set evaluation and systematic ablations remain future work.
- The framework combines masked-autoencoder pre-training, tracer-specific segmentation models, anatomical organ guidance, and scribble-conditioned refinement.
- Ten-fold cross-validation improves mean Dice from 0.5889 to 0.6175 for FDG and from 0.5833 to 0.6405 for PSMA after refinement.
- Official AutoPET V test-set evaluation and systematic component ablations remain future work.