DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation

TL;DR

DyCON tackles the challenge of semi-supervised medical image segmentation under extreme class imbalance and pathology-driven uncertainty by introducing two specialized losses: UnCL, which dynamically weights voxel-wise consistency by entropy to favor uncertain regions early and refine confident areas later, and FeCL, which uses patch-wise focal, entropy-aware contrastive learning with top-$k$ hard negatives to improve local discrimination. The framework is embedded in a Mean-Teacher setup and leverages entropy-based weighting $\frac{1}{\exp(\beta \cdot H_s(p^s_i)) + \exp(\beta \cdot H_t(p^t_i))}$ with a schedulable $\beta(t)$ to balance exploration and convergence, plus patch-level representations $z_i^s$ processed through a projection head with dual focal weights $F_k^+$ and $F_q^-$. Extensive experiments on ISLES'22, BraTS'19, LA, and Pancreas demonstrate state-of-the-art performance under 5–20% labeled data, with notable improvements in Dice, IoU, and boundary delineation, confirming robust generalization across MRI and CT modalities. The results highlight the method’s potential to reduce annotation burden while preserving high segmentation fidelity in clinically challenging scenarios.

Abstract

Semi-supervised learning in medical image segmentation leverages unlabeled data to reduce annotation burdens through consistency learning. However, current methods struggle with class imbalance and high uncertainty from pathology variations, leading to inaccurate segmentation in 3D medical images. To address these challenges, we present DyCON, a Dynamic Uncertainty-aware Consistency and Contrastive Learning framework that enhances the generalization of consistency methods with two complementary losses: Uncertainty-aware Consistency Loss (UnCL) and Focal Entropy-aware Contrastive Loss (FeCL). UnCL enforces global consistency by dynamically weighting the contribution of each voxel to the consistency loss based on its uncertainty, preserving high-uncertainty regions instead of filtering them out. Initially, UnCL prioritizes learning from uncertain voxels with lower penalties, encouraging the model to explore challenging regions. As training progress, the penalty shift towards confident voxels to refine predictions and ensure global consistency. Meanwhile, FeCL enhances local feature discrimination in imbalanced regions by introducing dual focal mechanisms and adaptive confidence adjustments into the contrastive principle. These mechanisms jointly prioritizes hard positives and negatives while focusing on uncertain sample pairs, effectively capturing subtle lesion variations under class imbalance. Extensive evaluations on four diverse medical image segmentation datasets (ISLES'22, BraTS'19, LA, Pancreas) show DyCON's superior performance against SOTA methods.

Figures (9)

• Figure 1: Comparative performance (average Dice (%) and HD95 on ISLES'22 and BraTS'19) when integrating DyCON into MT (Mean-Teacher) and CT (Co-Training) frameworks with 10% labeled data. DyCON improves segmentation accuracy without modifying training pipelines, proving its versatile effectiveness.
• Figure 2: The overall framework of DyCON. Following the MT paradigm, the teacher model produces logits $p^t$ from unlabeled input $x_j^t$, while the student model generates logits $p^s$ and masks $p_i^s$ from unlabeled $x_j^s$ and labeled $x_i$ inputs, respectively. DyCON is ultimately optimized using the labeled ($\mathcal{L}_{Dice}$ and $\mathcal{L}_{CE}$) and unlabeled ($\mathcal{L}_{\text{UnCL}}$ and $\mathcal{L}_{\text{FeCL}}$) losses in an end-to-end semi-supervised manner. The FeCL loss uses patch embeddings ($z^s$) from a projection head ($h_\phi$) and lesion-wise hard negatives $\mathbf{S}_{il}$ to promote discriminative feature learning.
• Figure 3: Evolution of uncertainty maps with UnCL. Entropy maps over iterations show decreasing uncertainty, focusing around lesion regions and aligning closely with GT as training progresses.
• Figure 4: Lesion and Tumor segmentation visualization on ISLES'22 and BraTS'19 datasets with 10% labeled data.
• Figure 5: Batch-wise mean gradient magnitudes for different $\beta$ values over time. Solid lines show smoothed trends, while transparent lines capture original gradient variability. $T$ refer to adaptively changing values of $\beta$.