Abstract
Profiling of Patient-Derived organoids is necessary for drug screening and precision medicine. This step requires accurate segmentation of three-dimensional cellular structures followed by protein readouts. While fully Convolutional Neural Networks are widely used in medical image segmentation, they struggle to capture long-range dependencies necessary for accurate segmentation. On the other hand, transformer models have shown promise in capturing long-range dependencies and self-similarities. Motivated by this, we present 3D-Organoid-SwinNet, a unique segmentation model explicitly designed for organoid semantic segmentation. We evaluated the performance of our technique using an Organoid dataset from four breast cancer subtypes. We demonstrated consistent top-tier performance in both the validation and testing phases, achieving a Dice score of 94.91 while reducing the number of parameters to 21 million. Our findings indicate that the proposed model offers a foundation for transformer-based models designed for high-content profiling of organoid models.