Semantic Class Distribution Learning for Debiasing Semi-Supervised Medical Image Segmentation
Authors: Yingxue Su, Yiheng Zhong, Keying Zhu, Zimu Zhang, Zhuoru Zhang, Yifang Wang, Yuxin Zhang, Xinyuan Zheng, +2 more
Organizations: Carnegie Mellon University, United States · Yale University, United States · Xi’an Jiaotong-Liverpool University, China · University College London, London, United Kingdom · Wuhan University, China
Medical image segmentation is critical for computer-aided diagnosis. However, dense pixel-level annotation is time-consuming and costly, and medical datasets often exhibit severe class imbalance. Such an imbalance causes minority structures to be overwhelmed by dominant classes in feature representations, hindering the learning of discriminative features and making reliable segmentation particularly challenging. To address this, we propose the Semantic Class Distribution Learning (SCDL) framework, a plug-and-play module that mitigates supervision and representation biases by learning structured class-conditional feature distributions. SCDL integrates Class Distribution Bidirectional Alignment (CDBA) to align embeddings with learnable class proxies and leverages Semantic Anchor Constraints (SAC) to guide proxies using labeled data. Experiments on the Synapse and AMOS datasets demonstrate that SCDL largely improves segmentation performance across both overall and class-level metrics, with particularly notable gains for several low-frequency organs. Our anonymous code is released at https://anonymous.4open.science/r/SCDL.
Figures & tables
Fig. 1: Comparison of imbalance-aware SSMIS paradigms: (a) Self-generated supervision uses model-derived unlabeled signals and may amplify head-class bias. (b) Loss reweighting and output calibration rebalance training but do not constrain class-conditional features. (c) SCDL debiases feature distributions via CDBA, distribution priors, and SAC anchoring.
Fig. 2: Overview of SCDL. (a) CDBA learns class proxy distributions and aligns them with token features to generate proxy-guided priors for downstream tasks. (b) SAC extracts semantic anchors from labeled regions and uses them to regularize the proxies, ensuring consistent class semantics across categories.
Methods
Average Dice of Each Class
Sp
RK
LK
Ga
Es
Li
St
Ao
IVC
PSV
PA
RAG
LAG
VNet (fully) [ 32 ]
90.2
91.9
90.7
38.3
30.9
94.8
75.6
79.1
81.4
62.1
48.5
48.9
58.0
CLD [ 13 ]
83.3
86.7
85.7
1.3
0.0
85.9
49.1
74.5
76.3
52.4
33.8
14.1
0.0
DHC [ 22 ]
72.5
55.2
49.9
60.3
50.0
87.8
31.5
67.0
54.5
38.7
27.8
19.4
7.8
SimiS [ 33 ]
83.3
90.8
85.8
9.2
0.0
85.6
55.0
73.6
71.7
50.4
34.0
0.0
16.6
SKCDF [ 24 ]
85.5
85.6
82.4
17.3
32.8
87.6
45.3
70.7
70.4
44.9
32.5
30.3
31.1
TABLE II: Per-class Dice scores (%) on the Synapse dataset with 20% labeled data.
Methods
Average Dice of Each Class
Sp
RK
LK
Ga
Es
Li
St
Ao
IVC
PA
RAG
LAG
Du
Bl
P/U
VNet (fully) [ 32 ]
92.2
92.2
93.3
65.5
70.3
95.3
82.4
91.4
85.0
74.9
58.6
58.1
65.6
64.4
58.3
CLD [ 13 ]
67.2
68.5
71.4
41.0
21.0
76.1
42.4
69.8
52.1
37.9
24.7
23.4
22.7
38.1
35.2
DHC [ 22 ]
73.3
63.8
62.6
40.3
34.4
81.8
47.2
69.0
56.9
45.5
30.5
16.8
31.8
49.2
27.7
SimiS [ 33 ]
77.4
72.5
68.7
32.1
14.7
86.6
46.3
74.6
54.2
41.6
24.4
17.9
21.9
47.9
28.2
SKCDF [ 24 ]
79.0
74.2
76.6
36.1
39.9
87.5
48.2
76.9
59.7
49.3
33.0
24.4
32.6
46.0
25.0
TABLE III: Per-class Dice scores (%) on the AMOS dataset with 5% labeled data.
Fig. 3: Qualitative comparison on the Synapse dataset. Representative segmentation results are shown on two abdominal CT slices with zoomed-in views of challenging regions. Compared with DHC, SKCDF, GA-MagicNet, and GA-CPS, SCDL-GA-CPS produces masks that better match the ground-truth annotations, especially for small and anatomically adjacent organs such as the portal and splenic veins, pancreas, and inferior vena cava. These results suggest that SCDL helps reduce inter-class confusion under imbalanced organ distributions.
Fig. 5: Per-class Dice improvement heatmap on Synapse and AMOS. Each cell represents the Dice improvement after adding SCDL to the corresponding baseline. Δ DHC, Δ GA-MagicNet, and Δ GA-CPS denote the gains over DHC, GA-MagicNet, and GA-CPS, respectively. Classes are ordered by foreground voxel frequency, and the head-tail split follows Fig. 4 . The results show that SCDL brings larger gains to many tail classes, especially when coupled with stronger imbalance-aware baselines.
School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, Shanghai, China · Department of Ultrasound, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China