Robust Organoid OCT Segmentation via Multi-View Graph Fusion under Sparse Annotations and Culture Variations

Optical coherence tomography (OCT), with its high resolution, deep penetration, and non-invasive characteristics, is crucial for three-dimensional (3D) imaging of organoids. However, speckle noise, low contrast, and signal disturbances arising from culture environments and drug addition to the culture medium often blur organoid boundaries and cause background mis-segmentation, significantly reducing segmentation accuracy. Although two-dimensional (2D) segmentation methods are efficient and flexible, they cannot capture inter-slice context, leading to incomplete intra-slice segmentation and inter-slice discontinuities. Conversely, while 3D methods can model inter-slice contexts through inherent 3D receptive field, they face high annotation requirements and computational costs. To overcome these limitations, we propose the Multi-View Inter-Slice Feature Fusion Network (MIFFNet). MIFFNet explicitly constructs inter-slice and inter-spatial graph connections between the target slice and adjacent slices, employing graph convolutions to fuse multi-view contextual information and form a quasi-3D receptive field that integrates both axial and non-axial contexts, thereby enhancing segmentation completeness, inter-slice continuity, and robustness. MIFFNet can be seamlessly integrated into existing 2D segmentation frameworks, enabling multi-slice joint modeling without requiring densely annotated continuous slices. Experimental results demonstrate that MIFFNet-enhanced 2D networks achieve Dice score improvements of over 6%, and reduce background mis-segmentation rates by more than 30% under drug-influenced scenarios. The MIFFNet-enhanced segmentation network consistently outperforms existing 2D, 3D, and inter-slice methods across multi-patient, multi-drug, and multi-cancer datasets, enhancing organoid segmentation accuracy and robustness under challenging imaging conditions and annotation scarcity.