Organoids closely mimic the structural and functional characteristics of native tissues, serving as valuable models for tumor research and drug screening. While their morphology reflects developmental states, traditional fluorescence labeling can disrupt structural integrity, and automated analysis in non-destructive bright-field imaging remains challenging due to complex backgrounds and morphological variability. To address these challenges, we propose a LGBP-Net, a robust deep learning framework designed for automated organoid segmentation and tracking. The model integrates complementary local texture and global contextual features through a hybrid CNN-Transformer encoder and employs a novel Learnable Gaussian Band-Pass Fusion (LGBP) module to adaptively merge multi-scale representations in the frequency domain. A Bidirectional Cross-scale Fusion (BCF) Block is further introduced to enhance decoder-level interaction between high-resolution details and semantic contexts, followed by a hierarchical progressive upsampling strategy. Extensive experiments on multiple datasets demonstrate the superiority of our method. On the challenging self-constructed mice bladder organoid dataset, LGBP-Net achieved a Dice coefficient of 85.64% and a Jaccard index of 76.73%, outperforming the state-of-the-art CMU-Net baseline by 3.95% and 4.55%, respectively. Furthermore, it attained a Dice score of 97.41% on the public brain organoid dataset, validating its generalization capability. LGBP-Net provides a highly accurate and non-invasive computational tool for large-scale organoid analysis, facilitating dynamic phenotyping and drug response evaluation without the need for fluorescent labeling. • A bright-field bladder cancer organoid dataset with complex backgrounds and time-lapse sequences. • LGBP-Fusion decomposes encoder features into bands and fuses local and global cues. • A bidirectional cross-scale fusion block refines details via attention and gating. • A hierarchical BCF decoder enables robust multi-scale organoid segmentation.
Zhang et al. (Fri,) studied this question.