cs.CVAug 18, 2026

MaLViL: Multi-axis Low-rank Vision-LSTM for Medical Image Segmentation

Authors: Afshin Bozorgpour, Sina Ghorbani Kolahi, Moein Heidari, Ilker Hacihaliloglu, Dorit Merhof

Organizations: Faculty of Informatics and Data Science, University of Regensburg, Germany · Independent Computer Science Researcher, Tehran, Iran · University of British Columbia, Vancouver, BC, Canada · Fraunhofer Institute for Digital Medicine MEVIS, Bremen, Germany

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Vision-LSTM (ViL) enables efficient global modeling, but its cost still scales with the number of spatial tokens, so existing segmenters confine ViL to a coarse bottleneck and lose fine anatomical detail. Rasterizing 2D features into a 1D sequence further breaks adjacency across the orthogonal scan axis. We propose MaLViL, a Multi-axis Low-rank Vision-LSTM network that extends ViL across decoder resolutions. Bidirectional low-rank ViL (Bi-LRViL) reasons on a compact orthonormal subspace and preserves detail through an orthogonal residual; scale-aware SaLViL restores cross-axis neighbors before serialization; and a Cross-Directional Mixer (CDM) fuses orthogonal horizontal and vertical traversal paths. Statistics-Guided Skip Modulation (SGSM) further retains boundary cues in encoder skips. On skin-lesion, ultrasound, and multi-organ CT benchmarks, MaLViL achieves competitive or state-of-the-art segmentation accuracy, while reducing ViL operator memory by up to 83×83\times at fine decoder resolutions. Code is available at: https://github.com/xmindflow/malvil.

Figures & tables

Figure 1: MaLViL architecture. A hierarchical encoder feeds four MaLViL Blocks; the three encoder skips are fused by SGSM. Each block contains a Cross-Directional Mixer (CDM) followed by a Spatial Quadratic FFN (SQ-FFN). The dashed path denotes the auxiliary low-rank reconstruction loss.
Figure 2: SaLViL and Bi-LRViL. Left: scale-specific branches recover orthogonal neighbors before Bi-LRViL sequence modeling. Right: rank- pp projection with opposite forward/backward ViL scans and gated fusion ( α\alpha , ω\omega ).
MethodsParamsFLOPsSpl.RKid.LKid.Gal.Liv.Sto.Aor.Pan.Average
DSC ↑\uparrowHD95 ↓\downarrow
R50 U-Net [ 8 ]30.42 M-85.8778.1980.6063.6693.7474.1687.7456.9074.6836.87
TransUNet [ 8 ]96.07 M88.91 G85.0877.0281.8763.1694.0875.6287.2355.8677.4931.69
Swin-UNet [ 6 ]27.17 M6.16 G90.6679.6183.2866.5394.2976.6085.4756.5879.1321.55
VM-UNet [ 20 ]44.27 M6.52 G89.5182.7686.1669.4194.1781.4086.4058.8081.0819.21
PVT-EMCAD-B2 [ 18 ]26.76 M5.60 G92.1784.1088.0868.8795.2683.9288.1468.5183.6315.68
Table 1: Evaluation results on the Synapse dataset ( blue indicates the best and red the second best results).
Figure 3: Qualitative comparison on the Synapse multi-organ dataset.
Methods PH 2 HAM10000 Dice Acc. Dice Acc. U-Net [ 19 ] 89.36 92.33 91.67 95.67 TransUNet [ 8 ] 88.40 92.00 93.53 96.49 Swin-Unet [ 6 ] 94.49 96.78 92.63 96.16 Att-UNet [ 17 ] 90.03 92.76 92.68 96.10 UCTransNet [ 22 ] 90.93 94.08 93.46 96.84 MissFormer [ 12 ] 85.50 90.50 92.11 96.21 CENet [ 4 ] 95.04 97.19 94.71 97.10 UxLSTM-Bot [ 9 ] 91.46 94.46 93.80 96.87 UxLSTM-Enc [ 9 ] 91.51 94.42 93.71 96.89 MaLViL 95.68 97.71 95.17 97.33Methods ISIC2017 ISIC2018 DSC ↑\uparrow ACC ↑\uparrow DSC ↑\uparrow ACC ↑\uparrow U-Net [ 19 ] 89.89 96.13 88.51 95.39 VM-UNet [ 20 ] 90.70 96.45 89.91 95.54 VM-UNet v2 [ 25 ] 90.45 96.37 89.02 95.51 DermoSegDiff [ 5 ] 91.43 96.72 89.66 95.75 EGE-UNet [ 21 ] 90.73 96.42 88.19 95.10 GLM-SFNet [ 7 ] 92.18 97.08 90.64 96.04 UltraLight VM-UNet [ 24 ] 90.91 96.46 89.40 95.58 UxLSTM-Bot [ 9 ] 90.00 96.46 89.92 96.02 UxLSTM-Enc [ 9 ] 90.19 96.38 89.78 96.04 MaLViL 92.08 97.15 91.23 96.75Methods BUSI Dice ↑\uparrow U-Net [ 19 ] 74.04 PraNet [ 10 ] 75.41 CaraNet [ 16 ] 77.34 Swin-UNet [ 6 ] 77.38 TranFuse [ 27 ] 79.36 U-Kan [ 15 ] 76.40 UWT-net [ 26 ] 79.65 UxLSTM-Bot [ 9 ] 72.48 UxLSTM-Enc [ 9 ] 72.80 MaLViL 81.76
Table 2: Comparison on PH 2 , HAM10000, ISIC’17, ISIC’18, and BUSI datasets.
Figure 4: Qualitative segmentation on PH 2 , HAM10000, and BUSI.
Figure 5: Internal features on ISIC 2017 ( 28×2828{\times}28 ): Bi-LRViL, SaLViL/CDM, and SGSM activations (normalized per map).
More in this topic

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 788 papers

Explore similar work

CardsList

  1. HSMLA: Hierarchical Softmax Multi-scale Linear Attention for Efficient Vision Transformers

    Aug 7, 2026Dong Liu, Yanxuan Yu, Renata Borovica-Gajic +2Efficient ViTsLinear Attention

    University of California, Los Angeles, USA · Columbia University, USA · The University of Melbourne, Australia +1

  2. LUMOS: Latent Universal Medical Priors for Segmentation

    Mar 1, 2026Zhuonan Liang, Wei Guo, Jie Gan +4Medical Image SegmentationTransfer Learning

    The University of Sydney, Sydney, NSW 2006, Australia. · Biological System & Engineering Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA. · Berkeley Biomedical Data Science Center, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA.

  3. LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder

    Sep 23, 2026Andrei Arhire, Mihaela-Elena Breabăn, Radu TimofteMedical Image SegmentationLightweight CNNs

    Faculty of Computer Science Alexandru Ioan Cuza University of Iasi · Computer Vision Lab, CAIDAS & IFI University of Wurzburg, Germany