The recent advancements in medical image segmentation have seen a significant shift towards leveraging multi-modal data and semi-supervised learning techniques to address the scarcity of labeled samples. Innovations in this area are focusing on improving the robustness and accuracy of segmentation models by incorporating global structural features and local details, often through the use of second-order statistical information. Additionally, there is a growing emphasis on developing models that can handle noisy annotations and inter-rater variability, which are common challenges in medical imaging. These models are designed to dynamically adjust learning weights and enforce consistency at both image and feature levels, ensuring balanced multi-modal learning. Notably, the integration of modality-specific and modality-invariant features is proving to be a key factor in enhancing the performance of these models. The results from recent studies indicate substantial improvements in segmentation accuracy, with some methods achieving up to a 10% gain in F1-score or MAPE, and demonstrating robust performance across diverse demographic factors and clinical tasks. These developments suggest a promising direction for future research in medical image analysis, potentially leading to more accurate and reliable diagnostic tools.