MI²-Net: A Mamba-based network for joint incomplete multi-modal and incomplete label MRI image segmentation.

Abstract

Current multi-modal segmentation methods typically rely on complete data and labels. However, clinical practice often faces the dual challenges of incomplete modalities and sparse annotations. Existing approaches mostly address these two issues in isolation, overlooking the compound impact of simultaneous deficiencies, thereby limiting their practical effectiveness. In this work, we propose MI²-Net, the first framework specifically designed for multi-modal MRI segmentation under the dual-missing setting. First, to tackle modality incompleteness, we introduce a missing-aware mamba imputation & fusion module. This module leverages a parallel tri-planar global-local mamba encoder for efficient intra-modal long-range representation learning. At the bottleneck, a multi-scale mamba imputation block is employed to reconstruct missing semantic features across different levels. Subsequently, a hybrid mamba fusion module is introduced to enhance inter-modal feature interactions. Second, to mitigate label scarcity, we employ a fusion-based semi-supervised learning strategy within a dual-decoder architecture, effectively propagating knowledge from labeled to unlabeled data via mutual consistency regularization. Furthermore, a specific semi-supervised auxiliary regularizer is designed to bolster feature representations. By utilizing high-quality soft pseudo-labels from the fusion branch as supervisory signals, this auxiliary task regularizes the modality-specific decoding process. Experimental results on three public datasets demonstrate that MI²-Net outperforms representative missing-modality and semi-supervised segmentation methods. Notably, despite addressing the more challenging dual-missing scenario, MI²-Net trained with 74 labeled and 73 unlabeled cases achieves an average DSC gain exceeding 0.90% over missing-modality segmentation methods trained with 147 labeled cases on BraTS2018. The code is available at https://github.com/ZHT-948/MI2-Net.