Semantic consistency-guided patch-wise relation graph reasoning scheme for lung cancer organoid segmentation in...

Patient-derived organoids (PDOs) have recently opened new possibilities for personalized precision medicine. Specifically, the segmentation of live PDOs in brightfield microscopy is essential for organoid morphological evaluation, cell viability assays, and drug screening. However, accurately and automatically annotating live organoids remains challenging due to factors such as low contrast, varying organoid characteristics, imaging artifacts, and the scarcity of large-scale annotated datasets. This study aims to address these challenges by proposing an advanced method for live PDOs segmentation even with limited training data.

We propose a semantic consistency-guided patch-wise relation graph reasoning scheme, which allows local-global semantics extraction, multi-task-based representation refinement and semantic consistency constraint for live PDOs segmentation. Specifically, our approach consists of two input-specific streams designed to extract high-level whole-image local and patch-wise semantic representations. Then a patch-wise relation graph convolution module (PRGCM) is devised to exploit global prior morphological properties by integrating low-level spatial relatedness into patch-wise high-level semantic representations using graph convolution. Subsequently, a local-global semantics fusion module (LGSFM) is deployed to enable local-global contextual semantic fusion. In the decoding stage, we develop two auxiliary learning tasks, with a patch segmentation decoder to guarantee semantic homogeneity by a novel loss function, called stream consistency (SC) loss, along with a reconstruction decoder to capture powerful and discriminative representation without additional annotations.

We introduce LiveOrganoid, a large-scale dataset of manually annotated lung cancer brightfield images, consisting of lung PDOs with diverse morphologies. Our scheme demonstrates superior performance on LiveOrganoid, achieving state-of-the-art results compared to recent methods, even with fewer training data. Additionally, we evaluate our method's generalization for different imaging modality and other cell-culture segmentation tasks, including cell and nuclear segmentation in tissue images.

Our proposed semantic consistency-guided patchwise relation graph reasoning scheme addresses the challenges in live PDOs segmentation, paving the way for improved organoid morphological evaluation, cell viability assays, and drug screening. The method's generalizability to other cell-culture segmentation tasks highlights its potential for broader applications in precision medicine.