LETT-NeXt: A Lightweight RECIST-Guided Model for 3D CT Lesion Segmentation

RECIST (Response Evaluation Criteria in Solid Tumors) is the clinical standard for assessing tumor response, but it relies on simple 2D diameter measurements that fail to capture the full 3D extent of a lesion. The CVPR 2026 Foundation Models for Pan-cancer Segmentation in CT Images competition challenged researchers to develop models that can predict accurate 3D lesion masks using these sparse RECIST markers as prompts.

LETT-NeXt addresses this by using a lightweight, RECIST-guided architecture. The model takes a 3D CT volume and a RECIST marker as input. It extracts a regional crop centered on the lesion, encodes the RECIST line and endpoints as additional input channels, and processes this data through a compact MedNeXt-v2 encoder-decoder. To handle cases where the initial crop might truncate the lesion, the model employs an adaptive inference strategy called AutoZoom, which expands the field of view when necessary.

LETT-NeXt demonstrates that sparse RECIST markers are sufficient to guide accurate 3D lesion segmentation under strict computational constraints. On the competition's hidden test set, the model achieved a Dice Similarity Coefficient (DSC) of 73.9 and a Normalized Surface Dice (NSD) of 67.3. Crucially, the model is highly efficient, completing inference on a CPU in under 7 seconds per case with a peak memory usage of approximately 3.6 GB.

Controlled ablation studies revealed that the choice of the MedNeXt-v2 backbone was the primary driver of performance, significantly outperforming other architectures like SwinUNETR. While the inclusion of an auxiliary anatomy-tumor head provided a minor boost, it acted more as a training-time regularizer than a transformative feature. AutoZoom proved to be a useful fallback mechanism for edge cases, though it had a neutral impact on average performance.

This work bridges the gap between clinical practice and volumetric analysis. By enabling 3D segmentation from the same RECIST markers already used in routine clinical workflows, LETT-NeXt allows for more comprehensive tumor burden assessment without requiring time-consuming manual 3D annotations. Its lightweight design ensures that such advanced analysis can be deployed in resource-constrained clinical environments without requiring high-end GPU hardware.