Abstract
In this study, we investigate the application of deep learning techniques for automatic segmentation of the Fluo-N2DH-GOWT1 dataset, which consists of time-lapse grayscale images of rat neural stem cells. We employ the DeepLabV3+ semantic segmentation framework with a ResNet-18 backbone, chosen for its balance between accuracy and computational efficiency on relatively small biomedical datasets. To improve generalization and robustness, we apply data augmentation strategies including rotation, scaling, shear, and reflection. The performance of the proposed model is evaluated using standard metrics such as F1-score, Intersection of Union, Precision and Recall. Experimental results demonstrate that the ResNet-18–based DeepLabV3+ achieves reliable segmentation of stem cells, effectively distinguishing cells from the background.