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
cs.CV · cs.AI
Quick verdict
TRIAGE: MAE-pretrained STU-Net + organ priors + scribble refinement for AutoPET V; gains shown on CV only, test results placeholders, key parts lack ablations.
Summary
The paper presents TRIAGE, a pipeline built for the AutoPET V challenge on interactive whole-body lesion segmentation in FDG and PSMA PET/CT. A 3D STU-Net-Small backbone is initialized by masked-autoencoder pre-training in which PET and CT are masked independently at a ratio of 0.5. A tracer classifier routes each study to one of two identically structured but independently trained FDG/PSMA branches; each branch uses a first-stage model (PET, CT, and organ masks from an auxiliary organ segmentation network) to produce an initial lesion mask, then a second-stage model refines it using the initial prediction and cumulative foreground/background scribbles, with curriculum-style training and multi-fold ensembling, followed by tracer-specific SUV thresholding (1.5 FDG / 1.0 PSMA g/mL). On official training data (1,014 FDG studies/900 patients; 597 PSMA/378 patients), ten-fold cross-validation yields stage-1 Dice of 0.5889 (FDG) and 0.5833 (PSMA), improving to 0.6175 and 0.6405 after refinement. A single FDG-fold ablation attributes +0.0999 Dice to pre-training. Main limitations: no test-set results (the authors explicitly call their numbers placeholders pending challenge evaluation); no ablations of organ context, curriculum, ensembling, routing, or SUV post-processing; an unspecified scribble simulation protocol for stage-2 validation; no significance testing; and FDG fold-4 stage-2 training left incomplete and excluded. Code is released on GitHub.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Asynchronous MAE pre-training (independent PET/CT masks, ratio 0.5) gives +0.0999 validation Dice over training STU-Net-Small from scratch in a single FDG-fold ablation without organ context (0.3890 to 0.4889).
- Two-stage interactive refinement raises ten-fold CV Dice from 0.5889 to 0.6175 (FDG, 9 folds) and 0.5833 to 0.6405 (PSMA).
- Tracer-specific FDG/PSMA branches selected by a routing classifier handle tracer-specific uptake better than a single shared model would.
- Anatomical organ context from an auxiliary segmentation model helps separate physiological uptake from malignant lesions.
- Curriculum-style training over interaction difficulty and fold ensembling improve robustness across interaction steps (asserted, not measured).
Strengths
Well-motivated tracer-aware decomposition
Secs. 1-2 argue FDG (Tübingen) and PSMA (LMU) cohorts differ in physiological uptake and lesion appearance; independently trained branches on 1,014 FDG vs 597 PSMA studies address this heterogeneity rather than forcing a shared model.
Pre-training ablation shows a large gain at zero inference cost
Table 3: STU-Net+MAE reaches 0.4889 Dice vs 0.3890 from scratch (+0.0999) at identical 14.55M parameters and 138.70 GFLOPs, while nnU-Net needs 30.79M/526.29 GFLOPs for 0.3990 (single FDG fold, no organ context).
Consistent two-stage improvement across tracers
Transparency about incompleteness
End-to-end reproducibility scaffolding
Weaknesses
No test-set or final challenge results
Most headline components lack ablation
Interactive evaluation protocol unspecified
No statistical testing and asymmetric fold reporting
Several design details too thin to reproduce
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TRIAGE is a challenge-method paper for AutoPET V (interactive whole-body PET/CT lesion segmentation, FDG + PSMA) from an NKI/Radboud/Amsterdam UMC/Macao Polytechnic team. Sec. 1 states three contributions: MAE-pretrained STU-Net features combined with organ-level anatomical context; interactive refinement formulated as second-stage segmentation conditioned on image channels, the initial prediction, and cumulative scribbles; and curriculum training over interaction difficulty plus fold ensembling. Critically, the abstract declares that quantitative results, ablations, and final test-set performance are left as placeholders to be completed after the challenge evaluation — every number in the paper is internal cross-validation on training data.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the TRIAGE pipeline for AutoPET V?
TRIAGE (Tracer-aware Refinement via Interactive Anatomy-Guided sEgmentation) is a two-stage interactive lesion segmentation system for whole-body PET/CT built for the AutoPET V challenge. A tracer classifier routes each study to an FDG- or PSMA-specific branch; within each branch a first-stage 3D STU-Net produces an initial lesion mask from PET, CT, and predicted organ masks, and a second-stage network refines it using the initial prediction plus user-drawn foreground/background scribbles. The backbone is initialized with masked-autoencoder pre-training in which PET and CT are masked independently.
What Dice scores did TRIAGE achieve?
On the official AutoPET V training data, ten-fold cross-validation gave a mean validation Dice of 0.5889 (FDG) and 0.5833 (PSMA) for the initial automatic segmentation, improving to 0.6175 (FDG) and 0.6405 (PSMA) with the interactive refinement stage. A single FDG-fold ablation showed pre-training raised Dice from 0.3890 to 0.4889. Test-set results were not yet available at publication; the authors explicitly label their reported numbers as placeholders pending the challenge evaluation.
Is the code and data for this paper available?
Code is released at github.com/Liiiii2101/AUTOPET2026-MEDAI. The study used only the official AutoPET V training dataset (1,014 FDG studies from 900 patients and 597 PSMA studies from 378 patients) with no external imaging data; the dataset is distributed through the challenge organizers. Training ran on a single NVIDIA RTX A6000.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 is present by caption only; the diagram itself was not extracted, so visual pipeline details could not be checked.
- Table 4 contains literal template artifacts (e.g. 'standardized framework?'), an apparent unfilled cell in the source table rather than extraction loss.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.