Comprehensive benchmarking of deep learning approaches for automated astrocyte segmentation in traumatic brain injury

Astrocyte morphological changes and GFAP upregulation are hallmarks of traumatic brain injury (TBI) and quantifying these alterations in tissues is essential for assessing TBI severity and progression. However, conventional segmentation methods such as manual labeling or thresholding are labor-intensive and prone to artifacts. We systematically evaluated six deep learning segmentation architectures (U-Net, U-Net++, FPN, MANet, LinkNet, and PSPNet) paired with seven common backbones (ResNet50/101/152, MobileNetV2, VGG16/19, and EfficientNet-b4) for automated astrocyte segmentation. The performance was compared to segmentation tools such as Ilastik, Cellpose, and GESUnet. A dataset of 220 manually labeled GFAP-stained ferret brain images (631×486 pixels) was used, including 182 training images from a single TBI case and 38 test images from 18 ferrets under sham and TBI conditions. Models were trained with a learning rate of 0.0001 over 200 epochs and evaluated using metrics including Dice coefficient (DC or F1 Score), intersection over union (IoU), precision, accuracy, specificity, and sensitivity (or recall). Performance evaluation was conducted in two steps: 5-fold cross-validation and testing on a separate dataset. UNet++/VGG19 achieved the best results (IoU: 50.01%, DC: 65.48%), outperforming other configurations and existing tools (IoU: 28.00-39.86%, DC: 42.94-54.40%). This model demonstrated robust segmentation of complex astrocytic morphologies and enabled accurate quantification of astrocyte reactivity across experimental conditions, supporting the use of deep learning for automated pathology assessment in TBI.