Anomaly and Spatially Aware Pretraining for Brain Hemorrhage Segmentation in Limited Annotated Data Setting

Abstract

Lesion segmentation in medical imaging is challenging due to limited voxel-level annotations and high variability in pathological appearance. Although self-supervised learning (SSL) enables the use of unlabeled data, most contrastive approaches do not explicitly incorporate pathology-related information, and reconstruction-based methods often remain more effective for downstream segmentation tasks. We propose an anomaly-modulated spatially-aware contrastive learning framework (AMSA) that integrates anatomical localization with anomaly-derived signals to improve representation learning in low-label regimes. First, images are affinely registered and decomposed into spatially aligned patches. Next, approximate anomaly maps are obtained via an image-to-image translation model and summarized into continuous anomaly scores. Spatial distance between patches is used to weigh interactions in the contrastive objective. Additionally, the contrastive objective is modulated with an estimated anomaly burden, encouraging representations to remain anatomically consistent while incorporating pathology-related information into the representation space. The pretrained encoder is subsequently fine-tuned for segmentation. Experiments on brain CT hemorrhage segmentation show consistent improvements over fully supervised training and spatially aware contrastive pretraining, while achieving performance comparable to reconstruction-based pretraining methods.