TriGAN-SiaMT: A triple-segmentor adversarial network with bounding box priors for semi-supervised brain lesion segmentation

Abstract

Accurate brain lesion segmentation in MRI is critical for clinical decision-making, but pixel-wise annotations remain costly and time-consuming. We propose TriGAN-SiaMT, a novel semi-supervised segmentation framework that combines adversarial learning, consistency regularization, and bounding box priors. Our architecture comprises three segmentors (S₀, S₁, S₂) and two discriminators (D₀, D₁). It includes: (1) a supervised branch (S₀↔D₀) trained on a small labeled subset; (2) a Siamese branch (S₁↔D₁) with an identical architecture to S₀↔D₀, but trained on unlabeled data; and (3) a teacher branch (S₂) updated via exponential moving average (EMA) from S₁, following the Mean Teacher (MT) paradigm. The teacher S₂ generates pseudo-labels to supervise S₁. It also provides soft segmentations to guide D₁, which does not see any labeled data. The model enforces consistency at multiple levels: between S₀ and S₁ (Siamese consistency), and between S₁ and S₂ (EMA consistency). Bounding box priors are incorporated as weak supervision for both labeled and unlabeled images, improving lesion localization. Evaluated on the ISLES 2022 and BraTS 2019 datasets, TriGAN-SiaMT achieves DSC scores of 84.80 % and 86.32 %, respectively, using only 5 % labeled data. These results demonstrate strong performance under limited supervision and robust generalization across brain lesions.