L2GNet: Optimal Local-to-Global Representation of Anatomical Structures for Generalized Medical Image Segmentation

Vandan Gorade, Rekha Singhal, Neethi Dasu, Sparsh Mittal, KC Santosh, Debesh Jha
Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare, PMLR 317:110-116, 2026.

Abstract

Continuous Latent Space (CLS) and Discrete Latent Space (DLS) models, like AttnUNet and VQUNet, have excelled in medical image segmentation. In contrast, Synergistic Continuous and Discrete Latent Space (CDLS) models show promise in handling fine and coarse-grained information. However, they struggle with modeling long-range dependencies. CLS or CDLS-based models, such as TransUNet or SynergyNet are adept at capturing long-range dependencies. Since they rely heavily on feature pooling or aggregation using self-attention, they may capture dependencies among redundant regions. This hinders comprehension of anatomical structure content, poses challenges in modeling intra-class and inter-class dependencies, increases false negatives and compromises generalization. Addressing these issues, we propose L2GNet, which learns global dependencies by relating discrete codes obtained from DLS using optimal transport and aligning codes on a trainable reference. L2GNet achieves discriminative on-the-fly representation learning without an additional weight matrix in self-attention models, making it computationally efficient for medical applications. Extensive experiments on multiorgan segmentation and cardiac datasets demonstrate L2GNet’s superiority over state-of-the-art methods, including the CDLS method SynergyNet, offering a novel approach to enhance deep learning models’ performance in medical image analysis.

Cite this Paper

BibTeX
@InProceedings{pmlr-v317-gorade26a, title = {L2GNet: Optimal Local-to-Global Representation of Anatomical Structures for Generalized Medical Image Segmentation}, author = {Gorade, Vandan and Singhal, Rekha and Dasu, Neethi and Mittal, Sparsh and Santosh, KC and Jha, Debesh}, booktitle = {Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare}, pages = {110--116}, year = {2026}, editor = {Wu, Junde and Pan, Jiazhen and Zhu, Jiayuan and Luo, Luyang and Li, Yitong and Xu, Min and Jin, Yueming and Rueckert, Daniel}, volume = {317}, series = {Proceedings of Machine Learning Research}, month = {20--21 Jan}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v317/main/assets/gorade26a/gorade26a.pdf}, url = {https://proceedings.mlr.press/v317/gorade26a.html}, abstract = {Continuous Latent Space (CLS) and Discrete Latent Space (DLS) models, like AttnUNet and VQUNet, have excelled in medical image segmentation. In contrast, Synergistic Continuous and Discrete Latent Space (CDLS) models show promise in handling fine and coarse-grained information. However, they struggle with modeling long-range dependencies. CLS or CDLS-based models, such as TransUNet or SynergyNet are adept at capturing long-range dependencies. Since they rely heavily on feature pooling or aggregation using self-attention, they may capture dependencies among redundant regions. This hinders comprehension of anatomical structure content, poses challenges in modeling intra-class and inter-class dependencies, increases false negatives and compromises generalization. Addressing these issues, we propose L2GNet, which learns global dependencies by relating discrete codes obtained from DLS using optimal transport and aligning codes on a trainable reference. L2GNet achieves discriminative on-the-fly representation learning without an additional weight matrix in self-attention models, making it computationally efficient for medical applications. Extensive experiments on multiorgan segmentation and cardiac datasets demonstrate L2GNet’s superiority over state-of-the-art methods, including the CDLS method SynergyNet, offering a novel approach to enhance deep learning models’ performance in medical image analysis.} }
Endnote
%0 Conference Paper %T L2GNet: Optimal Local-to-Global Representation of Anatomical Structures for Generalized Medical Image Segmentation %A Vandan Gorade %A Rekha Singhal %A Neethi Dasu %A Sparsh Mittal %A KC Santosh %A Debesh Jha %B Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare %C Proceedings of Machine Learning Research %D 2026 %E Junde Wu %E Jiazhen Pan %E Jiayuan Zhu %E Luyang Luo %E Yitong Li %E Min Xu %E Yueming Jin %E Daniel Rueckert %F pmlr-v317-gorade26a %I PMLR %P 110--116 %U https://proceedings.mlr.press/v317/gorade26a.html %V 317 %X Continuous Latent Space (CLS) and Discrete Latent Space (DLS) models, like AttnUNet and VQUNet, have excelled in medical image segmentation. In contrast, Synergistic Continuous and Discrete Latent Space (CDLS) models show promise in handling fine and coarse-grained information. However, they struggle with modeling long-range dependencies. CLS or CDLS-based models, such as TransUNet or SynergyNet are adept at capturing long-range dependencies. Since they rely heavily on feature pooling or aggregation using self-attention, they may capture dependencies among redundant regions. This hinders comprehension of anatomical structure content, poses challenges in modeling intra-class and inter-class dependencies, increases false negatives and compromises generalization. Addressing these issues, we propose L2GNet, which learns global dependencies by relating discrete codes obtained from DLS using optimal transport and aligning codes on a trainable reference. L2GNet achieves discriminative on-the-fly representation learning without an additional weight matrix in self-attention models, making it computationally efficient for medical applications. Extensive experiments on multiorgan segmentation and cardiac datasets demonstrate L2GNet’s superiority over state-of-the-art methods, including the CDLS method SynergyNet, offering a novel approach to enhance deep learning models’ performance in medical image analysis.
APA
Gorade, V., Singhal, R., Dasu, N., Mittal, S., Santosh, K. & Jha, D.. (2026). L2GNet: Optimal Local-to-Global Representation of Anatomical Structures for Generalized Medical Image Segmentation. Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare, in Proceedings of Machine Learning Research 317:110-116 Available from https://proceedings.mlr.press/v317/gorade26a.html.

Related Material