Search papers, labs, and topics across Lattice.
While promptable foundation models like SAM 3 rely on frozen session memory that plateaus after minimal interaction, LeCor converts clinician correction clicks into test-time training signals on case-specific adapters. The adapter weights are meta-learned so that a single gradient update on a touched CT slice generalizes segmentation improvements to untouched volumetric slices. Evaluated on 3D lung CT cohorts, this approach matches the 7-round correction accuracy of fine-tuned foundation models in just 3 rounds and increases final multi-slice Dice from 0.787 to 0.827.
In-context memory plateaus fast during interactive 3D segmentation, but turning clinician clicks into meta-learned test-time gradient steps halves human correction effort while cutting volumetric failure rates by over 40%.
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.
