Open Access

ARTICLE

Scale Ladder Consistency for Structure-Aware Multimodal Representation Learning in 3D Medical Image Segmentation

Weiqing Liu^1,#, Bin Li^1,#,*, Lianfang Tian¹, Qianhui Qiu²

1 School of Automation Science and Engineering, South China University of Technology, Guangzhou, China
2 Department of Otolaryngology, Head and Neck Surgery, Guangdong Provincial People’s Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China

* Corresponding Author: Bin Li. Email:
# These authors contributed equally to this work

(This article belongs to the Special Issue: Emerging Artificial Intelligence Technologies and Applications-II)

Computer Modeling in Engineering & Sciences 2026, 148(2), 39 https://doi.org/10.32604/cmes.2026.087647

Received 20 June 2026; Accepted 12 August 2026; Issue published 28 August 2026

Abstract

Self-supervised representation learning can reduce the dependence of three-dimensional (3D) medical image segmentation on dense voxel annotations. In multimodal 3D medical imaging, intensity-reconstruction pre-training provides dense appearance supervision but does not explicitly distinguish the structural regions that determine segmentation boundaries and small targets. A second mismatch arises in scale learning: encoder-decoder networks provide multi-scale feature maps, but they do not explicitly supervise how fine anatomical structures weaken or persist across neighboring scales. To address these mismatches, this study proposes Scale Ladder Consistency (SLC), a structure-aware self-supervised representation learning framework for multimodal 3D medical image segmentation. SLC combines Scale-Space Structural Reconstruction (SSR), Hybrid Mask, and Scale Ladder (SL) in its structural pre-training path. SSR replaces intensity recovery with structure prediction, Hybrid Mask increases supervision on fine structural regions, and SL learns neighboring-scale structural transitions through bidirectional prediction. Subset-to-Full Regularization (S2F) further stabilizes case-level representations during pre-training. Experimental results on the Brain Tumor Segmentation 2019 (BraTS19) and carotid artery datasets demonstrate that SLC consistently outperforms matched scratch fine-tuning and achieves competitive performance against recent segmentation and self-supervised methods. These results indicate that structure-aware and cross-scale self-supervised objectives can provide effective representations for multimodal 3D medical segmentation.

---

Keywords

---