Comparative Evaluation of En Bloc Staining Protocols for Deep-Learning Segmentation of Dorsal Root Ganglia in SBF-SEM

The integration of deep learning algorithms into volumetric electron microscopy can be limited by biochemical sample preparation. En bloc staining protocols, emphasizing heavy-metal contrast, are optimized for the convenience of human visual interpretation. This study evaluates the suitability of three staining protocols for automated segmentation and 3D reconstruction of dorsal root ganglia (DRG): rOTO-based heavy-metal staining (Ellisman), standard transmission electron microscopy (TEM) staining (Palade), and uranyl-free en bloc staining (X-solution). U-Net-based deep learning models were applied to dorsal root ganglia (DRG) samples to segment nuclei, mitochondria, satellite glial cells (SGCs), and the endoplasmic reticulum (ER). Quantitative segmentation performance was evaluated using the Dice similarity coefficient (DSC) and intersection over union (IoU) for each segmented structure group based on the comparison between the automated prediction masks and 30 manually annotated ground truth (GT) sections per staining protocol. Although the Ellisman-stained samples produced strong ultrastructural contrast, more false-positive and false-negative regions were observed for several structures. The Palade protocol provided insufficient membrane contrast to reliably identify the ER in this dataset. In contrast, the uranyl-free protocol produced the highest DSC and IoU values for SGCs and mitochondria and higher ER segmentation scores than the Ellisman protocol.