Source-linked AI summary
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
TL;DR
Lung-tumour contouring is time-consuming and variable, motivating interactive segmentation with clinician corrections. LeCor meta-learns test-time updates to case adapters so each click improves unclicked slices. On held-out lesions spanning at least eight slices, it improves seven-round Dice and reaches the fine-tuned model’s seven-round accuracy in three rounds.
Problem
Manual lung-tumour delineation is time-consuming and variable, while fixed-weight memory conditioning stops improving untouched slices after repeated corrections.
Method
LeCor uses case adapters reset per case and meta-learns their initialisation and update steps so click-driven gradients improve unclicked slices.
Results
4.00 points of Dice separate LeCor from the fine-tuned model after seven rounds, and three rounds reach the fine-tuned model’s seven-round accuracy.
Takeaways & Limitations
LeCor reaches moderate contour quality in fewer correction rounds and reduces failures at the Dice ≥0.80 target from 47 to 27 cases.
Takeaways & Limitations
The results are established for lesions spanning at least eight slices, and simulated clicks always target the deepest current error rather than clinician-chosen locations.
Abstract
from arXiv · showhide
Delineating lung tumours on computed tomography (CT) takes a considerable share of the time spent on radiotherapy planning, and a contour proposed by a model can be refined interactively by the clinician. Promptable foundation models such as SAM 3 support this workflow by writing each correction into a session memory that conditions the remaining slices, while the model weights stay fixed. On 690 test cases from five public CT cohorts, fine-tuning SAM 3 on lung tumours raises the Dice obtained from a single point prompt from 0.298 to 0.757, and seven rounds of corrections raise it further to 0.765, but under memory conditioning alone the accuracy on slices the annotator has not touched stops improving after six rounds. We therefore treat each correction as a training signal and propose LeCor, which performs test-time training on a small set of case adapters that are reset for every case and meta-learned such that a single gradient step driven by a click improves the slices that were not clicked. On the 133 test cases that span at least eight slices, LeCor raises the Dice reached after seven correction rounds from 0.787 with the fine-tuned model to 0.827, reduces the number of cases that never reach a Dice of 0.80 from 47 to 27, and reaches in three correction rounds the accuracy that the fine-tuned model attains in seven.
1 Introduction
Interactive lung-tumour segmentation reduces the burden and variability of manual CT contouring, but fixed-weight memory conditioning stops improving untouched slices. LeCor instead meta-learns test-time updates so each correction improves the whole lesion.
- Motivation: 16 to 22 minutes are required for manual lung-tumour delineation per case, with about 1 cm three-dimensional inter-observer variation at one standard deviation.Interactive model proposals can make editing faster and more consistent than manual delineation.
- Motivation: Interactive workflows let clinicians refine model-proposed contours with corrective clicks instead of manually delineating every tumour contour.This workflow is motivated by evidence that editing proposals is faster and more consistent than manual delineation.
- Background: SAM-style 3D medical-image segmentation treats CT slices as time, storing clicks in session memory so subsequent slices attend to earlier corrections.The model weights remain fixed during this memory-conditioning process.
- Background: Existing test-time training methods use a fixed, hand-designed optimisation procedure, whereas LeCor learns how correction-driven updates should improve the remaining lesion.Interactive segmentation supplies user corrections as supervision, making test-time optimisation especially suitable for this setting.
- Contribution: LeCor meta-learns test-time updates so a click improves unclicked slices across the whole lesion, rather than only conditioning later slices through fixed model weights.Case adapters provide the changing parameters, while the update is trained using clicked slices and evaluated on unclicked slices.
- Contribution: 4.00 percentage points of Dice separate LeCor from the fine-tuned model after seven correction rounds, while three LeCor rounds reach the fine-tuned model’s seven-round accuracy.LeCor is also 2.82 points above the same update with a fixed step on 133 held-out lung tumours.
2 Related Work
Related work spans medical adaptations of SAM and interactive click protocols, as well as test-time optimisation and meta-learning. LeCor combines these directions by learning the update itself for interactive medical segmentation.
- Medical foundation-model adaptation: Medical foundation-model adaptations include encoder fine-tuning, LoRA, adapter blocks, convolutional low-rank branches, and video-model extensions across medical modalities.These methods adapt SAM-family models to medical images with varying parameter-efficiency strategies.
- Interactive segmentation: Interactive segmentation commonly simulates corrective clicks from prediction errors and evaluates clicks needed to reach target overlap together with cases that never reach it.RITM, FocalClick, and SimpleClick exemplify this protocol.
- Test-time optimisation: Some interactive methods optimise inputs or intermediate activations at click time, while others update network weights online from user corrections.These approaches differ in which variables are adapted while processing the interaction.
- LeCor positioning: LeCor’s overview contrasts fixed-weight memory conditioning with meta-training on clicked slices and inference updates that refine unclicked slices.The figure presents the mechanism connecting interactive segmentation, test-time training, and meta-learning.
- Test-time training and meta-learning: Test-time training adapts model weights on test inputs, while meta-learning methods learn initialisations for effective few-step adaptation and Meta-SGD additionally learns parameter-wise step sizes.This background motivates learning both how adaptation starts and how it proceeds.
3 Methods
LeCor combines a fine-tuned SAM 3 lung-tumour segmenter with resettable case adapters whose meta-learned updates use clicks to improve unclicked slices. Sessions apply one adapter update after each correction round, then propagate the updated model through the whole case.
- Interactive setup: Each case is processed as a z-ordered slice sequence with an initial middle-slice centroid point, followed by forward and backward propagation.The study uses five public CT collections and defines a case as a z-contiguous run of slices containing one lesion.
- Interactive setup: Corrective clicks are sampled from false-negative or false-positive pixels on at most four selected slices during training and across seven evaluation rounds.Clicks are positive on false negatives and negative on false positives, with occasional sampling from the reference mask.
- Learning to be corrected: LeCor changes only small case adapters during inference, resetting them per case while keeping the SAM 3 backbone and shared task adapters fixed.The case adapters occupy 0.09% of backbone parameters and are updated after each correction round.
- Test-time update: A correction drives one gradient step on the case adapters using click supervision and a consistency loss computed on clicked slices, with the session memory held fixed.The loss uses the annotator’s click labels and a model-generated pseudo-mask at inference.
- Meta-learning objective: Meta-training optimizes adapter initial values and per-tensor step sizes so updates reduce Dice and cross-entropy error on unclicked slices rather than merely fitting clicked slices.The outer objective evaluates later-round performance on the unclicked set and penalizes degradation before any click.
- Evaluation: Evaluation focuses on 133 test cases spanning at least eight slices, reporting unclicked-slice Dice across seven fixed correction rounds alongside segmentation and annotation-cost metrics.Reported metrics include Dice, IoU, precision, recall, HD95, NSD, rounds to target Dice, and cases failing to reach the target.
4 Results
On 690 cases, fine-tuning and simulated corrections establish a strong interactive SAM 3 baseline, while LeCor improves unclicked-slice accuracy and reduces correction effort through learned test-time updates.
- 4.1 Interactive base model: 0.298 to 0.757 Dice after fine-tuning SAM 3 with a single point prompt, and seven simulated correction rounds reach 0.765.These results are measured on the full 690-case test set and provide the interactive base model for subsequent experiments.
- 4.2 Accuracy across correction rounds: 0.827 Dice at round 7 with LeCor exceeds 0.791 without updates on unclicked slices across 133 cases.LeCor is still improving at round 7, whereas memory-only conditioning saturates by round 6 and declines at round 7.
- 4.3 Qualitative analysis: LeCor removes persistent contour leaks and recovers missed tumour regions on unclicked slices that memory-only conditioning leaves unresolved.Qualitative examples show LeCor reaching 0.91 on one slice versus 0.70 without updates and 0.79 with the fixed step.
- 4.4 Source of the gain: 2.24 percentage points come from meta-learned initial values over standard ones, while learned steps add 1.82 points when paired with those initial values.Learned steps applied to standard initial values are harmful, showing that the initialization and update rule must be matched.
- 4.5 Annotation cost: 27 cases with LeCor versus 47 without updates fail to reach Dice ≥0.80 within seven rounds.LeCor also exceeds the no-update best Dice from round 3, so three rounds achieve what memory conditioning reaches in seven.
- 4.6 Computational cost and portability: 24.6 seconds per case for LeCor versus 17.7 seconds without updates, with the additional cost attributed to a backward pass each round.The meta-learned parameters themselves do not account for the runtime increase.
5 Discussion
LeCor improves interactive lung-tumour segmentation through meta-learned case-adapter updates, reducing correction rounds and failures at moderate Dice targets. Its gains are concentrated at Dice ≥0.80, while the method remains limited to lesions spanning at least eight slices and simulated deepest-error clicks.
- Fixed correction rules improve segmentation only late and within a narrow step-size range, whereas meta-learning makes the same mechanism effective from the first correction.The controls implicate the learned initialization rather than the step rule as the source of the improvement.
- LeCor reaches moderate-quality contours in fewer rounds and substantially reduces cases whose corrections degrade the initial segmentation.The added computational cost is one backward pass per correction round.
- The reduction in rounds is established at Dice ≥0.80, not Dice ≥0.85, and evaluation covers lesions spanning at least eight slices with simulated deepest-error clicks.The meta-objective requires unclicked slices, and clinician-selected click behavior remains unmeasured.
6 Conclusion
LeCor trains an interactive lung-tumour segmentation model to learn from corrections through meta-learned updates of resettable case adapters. It improves Dice, reduces failures at Dice ≥0.80, and reaches seven-round memory-conditioning accuracy in three rounds, with one second of GPU time per round.
- LeCor gains 4.00 Dice points over the fine-tuned model at round 7 and 2.82 points over the same fixed-step update.The gain is produced by the test-time update and attributed to the meta-learned starting point.
- LeCor cuts failures at Dice ≥0.80 by 43% and reaches in three rounds what memory conditioning reaches in seven.The method uses a meta-learned update of a small set of case adapters.
- Training, evaluation, and analysis code will be released on acceptance, including scripts that regenerate the paper’s numbers and figures.