Organizations: School of Artificial Intelligence, Anhui University, 230601, Hefei, China · State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, Anhui University, Hefei, 230601, China · Anhui Provincial Key Laboratory of Security Artificial Intelligence, Anhui University, Hefei, 230601, China · Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, Anhui University, 230601, Hefei, China · School of Computer Science and Technology, Anhui University, 230601, Hefei, China
Medical image segmentation is crucial for computer-aided diagnosis, which necessitates understanding both coarse morphological and semantic structures, as well as carving fine boundaries. The morphological and semantic structures in medical images are beneficial and stable clues for target understanding. While the fine boundaries of medical targets (like tumors and lesions) are usually ambiguous and noisy since lesion overlap, annotation uncertainty, and so on, making it not reliable to serve as early supervision. However, existing methods simultaneously learn coarse structures and fine boundaries throughout the training process. In this paper, we propose a structure and progress-aware diffusion (SPAD) for medical image segmentation, which consists of a semantic-concentrated diffusion (ScD) and a boundary-centralized diffusion (BcD) modulated by a progress-aware scheduler (PaS). Specifically, the semantic-concentrated diffusion introduces anchor-preserved target perturbation, which perturbs pixels within a medical target but preserves unaltered areas as semantic anchors, encouraging the model to infer noisy target areas from the surrounding semantic context. The boundary-centralized diffusion introduces progress-aware boundary noise, which blurs unreliable and ambiguous boundaries, thus compelling the model to focus on coarse but stable anatomical morphology and global semantics. Furthermore, the progress-aware scheduler gradually modulates noise intensity of the ScD and BcD forming a coarse-to-fine diffusion paradigm, which encourage focusing on coarse morphological and semantic structures during early target understanding stages and gradually shifting to fine target boundaries during later contour adjusting stages.
Figures & tables
Figure 1: (a) Beneficial and stable morphological and semantic structures. (b) Ambiguous and unreliable medical boundaries. (c) Conventional training paradigm simultaneously learns coarse structures and fine boundaries throughout the whole training process. (d) The proposed coarse-to-fine paradigm prioritizes acquiring coarse and stable structures and gradually shifts to carving fine and unreliable boundaries in medical images.
Figure 2: Overview of the proposed structure and progress-aware diffusion (SPAD) framework. The conditional diffusion model progressively denoises the final segmentation prediction x0 from a random noise xT (initial noise input) and the perturbed segmentation image (progress-aware noise condition) step by step. The progress-aware noise condition is generated from the combination of boundary-centered perturbing and semantic-concentrated perturbing. At each step, the perturbing intensities are dynamically modulated by a progress-aware scheduler (PaS), enabling a progress-aware coordination of the coarse structures and fine boundaries in medical segmentation targets.
Figure 3: Illustration of the semantic-concentrated diffusion (ScD), which perturbs pixels within a specific medical target with progress-aware noise but preserves small unaltered areas as semantic anchors, encouraging structure-guided reasoning and context-based reconstruction.
Figure 4: Illustration of the boundary-centralized diffusion (BcD). It first extracts the target boundary with a contour detector (such as the Canny operator) from the ground-truth mask label, and then blurs the unreliable and ambiguous boundary with Gaussian noise to lower the reliance on uncertain boundaries during early target understanding stages.
Methods
mIoU (%) ↑
mDice (%) ↑
SRF
IRF
PED
SHRM
IS/OS
Mean
SRF
IRF
PED
SHRM
IS/OS
Mean
U-Net [ 4 ]
64.81
64.82
66.14
56.37
72.06
64.48
78.65
78.66
79.62
72.10
83.76
78.40
Swin-Unet [ 40 ]
69.92
58.28
64.20
62.75
75.65
65.93
82.30
73.64
78.20
77.11
86.14
79.47
TransUnet [ 41 ]
69.49
59.81
65.19
60.29
73.39
65.49
82.00
74.85
78.93
75.23
84.65
79.15
MSA [ 11 ]
55.07
45.23
54.94
55.93
49.58
52.03
71.03
62.29
70.92
71.74
66.29
68.45
D-LKA Net [ 14 ]
65.59
55.32
55.30
57.29
73.42
61.08
79.22
71.23
71.22
72.85
84.67
75.84
Table 1: Comparing with the state-of-the-art methods in medical image segmentation on the AMD-SD dataset
Methods
mIoU (%) ↑
mDice (%) ↑
U-Net [ 4 ]
66.85
77.35
Swin-Unet [ 40 ]
66.37
77.12
TransUNet [ 41 ]
67.12
77.53
MSA [ 11 ]
61.03
73.27
D-LKA Net [ 14 ]
69.53
79.54
DSC-AMP [ 15 ]
70.34
82.58
Table 2: Comparing with the state-of-the-art methods in medical image segmentation on the CXRS dataset
Methods
mIoU (%) ↑
mDice (%) ↑
Baseline [ 42 ]
69.79
82.20
ScD with PaS
70.93
83.00
BcD with PaS
71.04
83.06
ScD+BcD without PaS
44.07
61.17
ScD+BcD with PaS ( ours )
71.51
83.39
Table 3: Ablation results on the AMD-SD dataset
Figure 5: Qualitative visualization of segmentation results on the AMD-SD dataset. From bottom to top: ground truth (GT), segmentation mask predicted from our SPAD, the SPAD without ScD (SPAD w/o ScD), the SPAD without BcD (SPAD w/o BcD), and the diffusion baseline without both ScD and BcD (Baseline). (1) The white boxes in the left four columns indicate representative regions where the baseline model fails to correctly localize target structures or produces incorrect predictions. Compared with the baseline, the methods incorporating ScD exhibit improved structural localization in these regions. (2) The green boxes in the right four columns indicate boundary-sensitive regions where the baseline produces imprecise or irregular contours. Compared with the baseline, the methods incorporating BcD yield more accurate boundary delineation. (3) When both ScD and BcD are incorporated, the predictions combine the advantages of both components and achieve the best.
Methods
Training
Testing
UNet [ 4 ]
3.9h
0.02h
CCDM (diffusion baseline) [ 42 ]
21.5h
0.78h
SPAD (ours)
21.6h
0.78h
Table 4: Efficiency comparison
Figure 6: Analysis of the parameter sensitivity and training stability of the proposed SPAD. (a) Sensitivity analysis of the scheduler decay parameter β in the progress-aware scheduler (PaS). (b) Sensitivity analysis of the semantic anchor ratio r in the semantic-concentrated diffusion (ScD). (c) Convergence behavior of training the SPAD on the AMD-SD dataset.
Accurate medical image segmentation is fundamental to precision medicine, yet robust delineation remains challenging under heterogeneous appearances, ambiguous boundaries, and large anatomical variability. Similar intensity and texture patterns between targets and surrounding tissues often lead to blurred activations and unreliable separation. We attribute these failures to representation collapse during encoding and insufficient fine grained multi scale decoding. To address these issues, we propose Med DisSeg, a dispersion driven medical image segmentation framework that jointly improves representation learning and anatomical delineation. Med DisSeg combines a lightweight Dispersive Loss with adaptive attention for fine grained structure segmentation. The Dispersive Loss enlarges inter sample margins by treating in batch hidden representations as negative pairs, producing well dispersed and boundary aware embeddings with negligible overhead. Based on these enhanced representations, the encoder strengthens structure sensitive responses, while the decoder performs adaptive multi scale calibration to preserve complementary local texture and global shape information. Extensive experiments on five datasets spanning three imaging modalities demonstrate consistent state of the art performance. Moreover, Med DisSeg achieves competitive results on multi organ CT segmentation, supporting its robustness and cross task applicability.
Zhiquan Chen, Haitao Wang, Guowei Zou +1
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China
Acquiring pixel-level annotations for medical image segmentation is a severe bottleneck. Traditional U-Net architectures, while effective, learn local texture patterns and lack awareness of global anatomical structures, leading to boundary delineation failures in low-data regimes. This research paper proposes utilizing unsupervised Denoising Diffusion Probabilistic Models (DDPMs) to extract anatomical features. We train a DDPM on 21 unlabeled abdominal CT scans to learn structural representations, transferring the encoder weights to a downstream segmentation task evaluated on the BTCV multi-organ dataset. Diffusion pretraining significantly improved liver segmentation: Dice increased from 0.75±0.36 to 0.93±0.16 (p<5.33×10−26, 0.529 Cohen's d), Average Surface Distance (ASD) decreased by 66%, and 95th-percentile Hausdorff Distance (HD95) reduced by 45%. For kidney segmentation, Dice improved from 0.90±0.19 to 0.95±0.10 (p<4.01×10−11). Multi-organ pooled performance showed a 68% variance reduction and a 74% improvement in boundary precision (Dice 0.95±0.07). Crucially, frozen encoder models retained > 80% of fine-tuned performance without exposure to segmentation labels, proving the existence of learned anatomical priors. In low-data scenarios, diffusion-pretrained models maintained robust performance with only 50% (Dice: 0.92 liver, 0.94 kidney), 25%, and even 10% (Dice: 0.89 liver, 0.71 kidney) of labeled data. Using unlabeled images for diffusion-based pretraining successfully embeds robust anatomical features prior to human supervision, transforming U-Nets into anatomy-aware systems.
Akshat G, Divyansh Gupta, Shaleen Bhatnagar +2
Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal, India
Medical image segmentation remains challenging in low-data regimes, where scarce annotations often yield poor generalization and ambiguous boundaries with missing fine structures. Recent self-supervised pretraining has improved transferability, but it often exhibits a texture bias. In contrast, accurate segmentation is inherently geometry-aware and depends on both topological consistency and precise boundary preservation. To address this problem, we propose a two-stage framework that couples structure-aware encoder pretraining with boundary-oriented decoding. In Stage-1, we aim to learn structure-aware representations for downstream segmentation in low-data regimes. To this end, we propose Mixed-Domain MeanFlow Pretraining, which aligns images and binary masks in a shared latent space through latent transport regression, where masks act as conditional structural guidance rather than prediction targets, making the pretraining task-agnostic. To further improve training stability under scarce supervision, we incorporate a lightweight Dispersive Loss to prevent representation collapse. In Stage-2, we fine-tune the pretrained encoder with a lightweight decoder that combines Direct Attentional Fusion for adaptive cross-scale gating and Frequency-Directional Dynamic Convolution for high-frequency boundary refinement under appearance variation. Experiments on ISIC-2016, Kvasir-SEG, and GlaS demonstrate consistent gains over state-of-the-art methods, with improved robustness in low-data settings and sharper boundary delineation.
Zhiquan Chen, Haitao Wang, Guowei Zou +1
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China