Planning CT lung tumour contours still eat radiotherapy time. A model proposes a stack overlay, the clinician clicks the wrong slices, and most promptable trackers only write that click into a session memory while the weights stay frozen. LeCor treats each correction as a tiny training step on case adapters that reset per patient, so the click improves slices the annotator never touched. Yi Luo, Yike Guo, Wenxuan Li, Zongwei Zhou, Rui Zhang, and Kai Ding posted arXiv:2609.09477 on 8 September 2026 (Johns Hopkins and University of Minnesota). Code and models are promised on acceptance.
Viewer teams already sell the interactive hang: propose a GTV, let the oncologist correct a few slices, propagate through the volume. Memory-only SAM-style trackers help early, then stall. This paper is about making the correction itself a learned update, not another prompt token.
Where memory-only correction stops helping
Manual lung GTV delineation still runs roughly 16 to 22 minutes per case, and inter-observer spread on the order of a centimetre is a known radiotherapy headache. nnU-Net proposals still need review. SAM 3 can treat the CT Z axis like a short video, carry clicks in session memory, and revise untouched slices while the backbone stays fixed.
On 690 held-out tumours from five public CT cohorts, LoRA fine-tuning of SAM 3 lifts a single-point prompt from about 0.30 to about 0.76 Dice. Seven correction rounds under memory conditioning alone only nudge that to about 0.77 on the full test set, and on the 133 cases that span at least eight slices the no-update interactive baseline saturates around round 6 near 0.79 Dice on unclicked slices. After that, more clicks stop buying volume-wide gains. The click is consumed as conditioning. It never moves the parameters that produced the leak into consolidated lung or the pleural miss.
What LeCor actually does
LeCor keeps the SAM 3 interactive stack and splits adapters in two. Task adapters (LoRA) encode lung tumours on CT and stay fixed at inference. Case adapters are a smaller LoRA set (rank 4 on MLP projections of the last eight image-encoder blocks). They reset at the start of every case and are the only weights that change during a session.
A correction round places up to four clicks on the worst local errors, builds a supervised loss on those clicked slices (Dice plus cross-entropy against a target assembled from the clicks), and takes one gradient step on the case adapters with per-tensor learned step sizes. The volume is then re-propagated with the updated adapters. Case adapters accumulate across rounds inside a case and reset when the next patient opens.
The meta-training objective is the point of the paper. During training, after a simulated click step, the model is scored on the unclicked slices of the same volume. Initial values of the case adapters and their step sizes are meta-learned so one click-driven step improves slices that were never clicked. Fixed-step test-time training without that meta start is weak early and only slightly helpful late. Meta init without the update actually hurts. The gain is the starting point for the update, not extra capacity sitting idle.
Where it was measured
Training and evaluation draw from LIDC-IDRI, NSCLC-Radiomics, MSD Task06 Lung, LNDb, and 4D-Lung. The interactive base is measured on 690 3D test cases. LeCor comparisons focus on the 133 test cases with at least eight slices, because the meta-objective needs unclicked slices inside the same lesion.
After seven correction rounds on those multi-slice cases, LeCor reaches about 0.83 Dice on unclicked slices versus about 0.79 for the fine-tuned no-update model (about 4 points) and about 0.80 for the same update with a fixed step. Cases that never hit Dice 0.80 within seven rounds fall from 47 to 27. Three LeCor rounds match the accuracy the memory-only fine-tuned model needs seven rounds to reach. The paper reports about one second of GPU time per correction round. Metrics stay as support. The shop claim is fewer clicks to a usable GTV overlay on the untouched stack.
Limits are explicit. Results are for lesions spanning at least eight slices. Clicks in the study always target the deepest current error; clinician click habits may differ. Code is not public yet (promised on acceptance). Fail closed for tiny single-slice nodules under the current objective, for workflows that forbid any test-time gradient, and when you only need memory prompting without case-level adapters.
How this lands in a viewer
If you already hang planning CTs in a DICOM viewer, treat LeCor as an interactive refinement policy on top of a SAM 3 style tracker, not a new worklist. Load the volume, seed a middle-slice point, propagate the tumour overlay, accept clinician clicks as correction rounds, run one case-adapter step after each round, then rehang the editable GTV layer. Reset case adapters when the study changes. Keep task adapters and the meta-learned case init as one checkpoint. Budget roughly a second of GPU per round for the backward pass.
GazeRefine turns gaze into training-free prompts. This paper keeps the click workflow and teaches the model to learn from those clicks so untouched slices move with the corrected ones.
Rebuild from arXiv:2609.09477. As of 10 September 2026 the abstract and PDF respond (HTTP 200). Watch the authors’ acceptance note for the public code drop.
Sources
- Yi Luo, Yike Guo, Wenxuan Li, Zongwei Zhou, Rui Zhang, Kai Ding. LeCor: Learning to Be Corrected by Meta-Learned Test-Time Training for Interactive 3D Lung-Tumour Segmentation. arXiv:2609.09477, posted 8 September 2026. https://arxiv.org/abs/2609.09477 (HTTP 200 on 10 September 2026). PDF: https://arxiv.org/pdf/2609.09477 (HTTP 200 on 10 September 2026). Code: promised on acceptance.