LeCor: Learning to Be Corrected by Meta-Learned Test-Time Training for Interactive 3D Lung-Tumour Segmentation
Yi Luo, Yike Guo, Wenxuan Li, Zongwei Zhou, Rui Zhang, Kai Ding
cs.CV · cs.LG
Quick verdict
LeCor turns clicks into meta-learned test-time adapter updates for 3D lung tumour segmentation; solid ablations, but simulated clicks, no external baselines.
Summary
The paper proposes LeCor, an interactive 3D lung-tumour segmentation method that treats each clinician click as a gradient-based training signal rather than only a prompt. Built on a SAM 3 video tracker fine-tuned with LoRA on five public CT cohorts (7,321 lesion cases; 690 test cases), the method adds small per-case 'case adapters' (0.09% of backbone parameters) that are reset for every case and updated by one gradient step after each correction round. The adapters' initial values and per-tensor step sizes are meta-learned with a MAML-style outer objective placed on unclicked slices, so a click on at most four slices improves the rest of the lesion. On the 133 test cases spanning at least eight slices, Dice after seven rounds rises from 0.787 (no update) to 0.827, failures at Dice ≥ 0.80 fall from 47 to 27, and three rounds of LeCor match seven rounds of memory-only conditioning. Ablations attribute the gain to the meta-learned initialisation rather than added parameters or learned step sizes alone. Main limitations: evaluation restricted to multi-slice lesions with simulated deepest-error clicks (no clinician study), no head-to-head against prior click-time adaptation methods, task and case adapters must be released as one coupled checkpoint, and code is promised only on acceptance.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
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
- Fine-tuning SAM 3 with LoRA raises single-point-prompt Dice from 0.298 to 0.757 on 690 lung-tumour test cases, and seven rounds of memory-only correction add only about 0.8 points (0.765).
- Meta-learned test-time updates of per-case adapters raise round-7 Dice on unclicked slices from 0.787 to 0.827 on the 133 multi-slice test cases (+4.00 points, p<0.001).
- Three rounds of LeCor exceed the best value (0.791, round 6) that memory-only conditioning ever reaches.
- The gain is attributable to the meta-learned initialisation: added parameters alone are harmful, and learned step sizes help only from the meta-learned starting point.
- Failures at Dice ≥ 0.80 fall from 47 to 27 of 133 cases, at roughly one additional second of GPU time per correction round.
Strengths
Factorial ablation isolates the mechanism
Table 2 and Fig. 2b cross initialisation × step rule on the same 133 cases: added parameters alone hurt (−2.1), learned steps on standard init hurt (−1.3), meta-init + fixed step gives +2.2, meta-init + learned steps +4.0. Combined with the step sweep (≥0.02 harmful from standard init, 0.015–0.05 beneficial from meta-init), the 'initialisation, not optimiser' conclusion (§5) is genuinely earned.
Above-average statistical reporting
Per-case paired two-sided Wilcoxon signed-rank tests with explicit thresholds, 95% CIs from 10,000 bootstrap resamples, NoC/NoF with failures charged the full seven-round budget, and three Dice targets — including the honest admission that the rounds-saving at Dice ≥ 0.85 is not significant (Table 3).
Honest negative and portability results
Clinically framed cost accounting
Precise mechanistic framing
Weaknesses
No head-to-head with prior click-time optimisation methods
Evidence base is 133 of 690 test cases, biased toward larger lesions
Self-supervised consistency target at inference is unanalysed
Single seed, no run variance, no multiplicity control
Recall stagnates at 0.71 across all seven rounds (Table 1)
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
LeCor targets interactive 3D lung-tumour segmentation for radiotherapy planning. The central claim: promptable-model workflows (SAM 3 run as a video tracker over the slice axis) consume clinician clicks only as session-memory conditioning, so accuracy on unclicked slices saturates (best 0.791 at round 6, declining at round 7) while weights stay fixed. LeCor instead treats each correction round as one gradient step on small per-case LoRA case adapters (rank 4, 0.09% of backbone parameters, reset per case), whose initial values and 64 per-tensor step sizes are meta-learned (MAML + Meta-SGD style) with the outer objective evaluated on unclicked slices. Secondary claims: the gain comes from the meta-learned initialisation, not the extra parameters or the learned steps; three rounds of LeCor match seven rounds of memory-only conditioning; and failures at Dice ≥ 0.80 fall 43% (47→27) for +0.99 s per round.
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 LeCor and how does it differ from using SAM 3 interactively?
LeCor is an interactive 3D lung-tumour segmentation method. Ordinary SAM 3 usage writes each clinician click into a session memory while model weights stay fixed. LeCor additionally takes one gradient step on small per-case adapters after every correction round, and those adapters are meta-learned so that the step improves slices the clinician never clicked.
How much better is LeCor than the fine-tuned SAM 3 baseline?
On 133 test lung-tumour cases spanning at least eight CT slices, Dice after seven correction rounds is 0.827 versus 0.787 without the update; cases never reaching Dice 0.80 fall from 47 to 27; and three LeCor rounds reach what seven rounds of the baseline achieve.
Does LeCor slow down annotation or inference?
It adds one backward pass per correction round: 24.6 s versus 17.7 s per case on a single A100, about one extra second per round. In exchange it reaches moderate contour quality (Dice 0.80) in fewer correction rounds, reducing interaction effort.
Is the LeCor code and data available?
Training uses five public CT cohorts (LUNA25, NSCLC-Radiomics, MSD Task06 Lung, LNDb, 4D-Lung). The paper states code, models and analysis scripts will be released on acceptance; nothing is public yet, and the task and case adapters must be deployed together as one checkpoint.
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.
- Figures are extracted as loose text (axis labels, legend entries, embedded panel numbers), so the curve shapes in Fig. 2 and the qualitative panels of Fig. 3 cannot be assessed visually; every number cited in this review comes from Tables 1-3 or body text, which are intact.
- Section 3.2 reads 'native 10082 resolution of SAM 3', most plausibly an extraction artifact of '1008²'; it is not treated as a factual claim anywhere in this review.
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.