A similarity-aware network with contrastive optimization for biomedical image segmentation

Abstract

Accurate biomedical image segmentation is crucial for clinical diagnosis. Convolutional neural networks and Transformer-based models have been widely used for biomedical image segmentation and have improved segmentation accuracy across multiple imaging modalities. However, most existing methods rely heavily on supervised learning with large-scale annotated datasets, while professionally annotated biomedical images remain scarce. In addition, the geometric consistency available from augmented and unlabeled images is not always explicitly exploited. We proposed a similarity-aware network with contrastive optimization for biomedical image segmentation, termed SimBIS. SimBIS supplements pixel-wise supervised learning with an output-space consistency objective. Specifically, segmentation predictions generated from augmented views of the same object are projected into a seven-dimensional Hu-moment space, and the discrepancy between these projections is minimized during training. The Hu-moment is used as a parameter-free auxiliary descriptor of global mask geometry, which is invariant to translation, isotropic scaling, and in-plane rotation. Fine-grained and irregular boundaries are still optimized through the pixel-level supervised loss. SimBIS improved mDice and mIoU on both LID and SID datasets across multiple classic segmentation baselines. On publicly available biomedical segmentation tasks, including Kvasir-Seg, CVC-ClinicDB, and ISIC 2018, SimBIS not only improved the performance of the original backbone models but also outperformed other state-of-the-art methods on various metrics. Visual segmentation results further showed that SimBIS can effectively segment small targets and distinguish challenging structures, producing more accurate and consistent contours than existing models. SimBIS improves biomedical image segmentation by incorporating Hu-moment-based global shape consistency into end-to-end optimization while retaining pixel-wise supervision. It effectively leverages augmented and unlabeled data without introducing additional trainable parameters or inference-time cost, demonstrating strong potential for annotation-efficient biomedical image segmentation.