Slide
Centre Interdisciplinaire
de Recherche et d’Innovation
en Cybersécurité et Société
de Recherche et d’Innovation
en Cybersécurité et Société
1.
Allaoui, M. L.; Allili, M. S.
LG-TargetSeg: Deep semantic fusion via language-conditioned selective scanning for targeted medical image segmentation Article de journal
Dans: Information Fusion, vol. 139, 2027, ISSN: 15662535 (ISSN).
Résumé | Liens | BibTeX | Étiquettes: Computerized tomography, decoding, Deep vision-language fusion, Encoder-decoder, Gaussians, Medical image processing, Medical image segmentation, Modeling languages, Natural language processing systems, Reinforced gaussian dice loss, Selective scanning, Semantic fusion, Semantic segmentation, Semantics, Shallow levels, Signal encoding, State space methods, State-space modeling, State-space models, Training time
@article{allaouiLGTargetSegDeepSemantic2027,
title = {LG-TargetSeg: Deep semantic fusion via language-conditioned selective scanning for targeted medical image segmentation},
author = {M. L. Allaoui and M. S. Allili},
url = {https://www.scopus.com/pages/publications/105051242411?origin=resultslist},
doi = {10.1016/j.inffus.2026.104797},
issn = {15662535 (ISSN)},
year = {2027},
date = {2027-01-01},
journal = {Information Fusion},
volume = {139},
publisher = {Elsevier B.V.},
abstract = {Clinical language defines the semantic intent of targeted medical image segmentation, yet most vision–language methods incorporate it only at a shallow level or through training-time conditioning, leaving spatial reasoning largely language-blind. We propose LG-TargetSeg (Language-Guided Targeted Segmentation), a cross-modal framework that embeds clinical language directly into the selective scan of a Mamba-based encoder-decoder. Natural-language prompts, encoded by a frozen MedSigLIP foundation model, shape selective scanning throughout the encoder–decoder and additionally adapt the bottleneck state transition, alongside gated modulation and cross-attention, enabling semantic guidance to continuously shape feature propagation toward the target anatomy. A Text-Driven Selective Scan 2D (TD-SS2D) module achieves prompt-aware multi-scale fusion with linear complexity in the number of spatial tokens. To improve robustness to boundary ambiguity and class imbalance, we introduce the Reinforced Gaussian Dice Loss (RGDL) with learnable weighting parameters. Evaluated under both task-specific and universal generalization settings, LG-TargetSeg achieves state-of-the-art performance. Moreover, its prompt-driven selectivity remains reliable under genuine multi-target competition: on abdominal CT images containing up to 16 annotated organs, changing only the text prompt on a frozen model produces organ masks with less than 1% pairwise overlap. Additional materials and code are available at: https://github.com/MOHAMEDLamine77/LG-TargetSeg . © 2026 Elsevier B.V.},
keywords = {Computerized tomography, decoding, Deep vision-language fusion, Encoder-decoder, Gaussians, Medical image processing, Medical image segmentation, Modeling languages, Natural language processing systems, Reinforced gaussian dice loss, Selective scanning, Semantic fusion, Semantic segmentation, Semantics, Shallow levels, Signal encoding, State space methods, State-space modeling, State-space models, Training time},
pubstate = {published},
tppubtype = {article}
}
Clinical language defines the semantic intent of targeted medical image segmentation, yet most vision–language methods incorporate it only at a shallow level or through training-time conditioning, leaving spatial reasoning largely language-blind. We propose LG-TargetSeg (Language-Guided Targeted Segmentation), a cross-modal framework that embeds clinical language directly into the selective scan of a Mamba-based encoder-decoder. Natural-language prompts, encoded by a frozen MedSigLIP foundation model, shape selective scanning throughout the encoder–decoder and additionally adapt the bottleneck state transition, alongside gated modulation and cross-attention, enabling semantic guidance to continuously shape feature propagation toward the target anatomy. A Text-Driven Selective Scan 2D (TD-SS2D) module achieves prompt-aware multi-scale fusion with linear complexity in the number of spatial tokens. To improve robustness to boundary ambiguity and class imbalance, we introduce the Reinforced Gaussian Dice Loss (RGDL) with learnable weighting parameters. Evaluated under both task-specific and universal generalization settings, LG-TargetSeg achieves state-of-the-art performance. Moreover, its prompt-driven selectivity remains reliable under genuine multi-target competition: on abdominal CT images containing up to 16 annotated organs, changing only the text prompt on a frozen model produces organ masks with less than 1% pairwise overlap. Additional materials and code are available at: https://github.com/MOHAMEDLamine77/LG-TargetSeg . © 2026 Elsevier B.V.