A generalizable deep learning model for automated 3D segmentation of orthopteran head anatomy in micro-CT

Deep learning tools are increasingly used today, particularly in medical segmentation. A gap nonetheless remains in automating segmentation for insects. This work addresses the following question: can a generalist segmentation model, trained on several phylogenetically related orthopteran species, reliably automate head tissue segmentation from micro-CT images? To answer this, we used nnU-Net, a self-configuring 3D deep learning segmentation framework originally developed for medical imaging, whose core function, learning to recognize tissues of interest, applies directly to this context. Six anatomical classes were automated, comparing two training strategies: sequential fine-tuning, which adds species one at a time under the assumption that progressive learning would strengthen predictive power, and from-scratch training, in which the model learns the entire dataset simultaneously. The fine-tuning model (ModelB) reached a Dice coefficient (a measure of overlap between automated segmentation and manual ground truth, ranging from 0 to 1) of 0.7715, compared to 0.7664 for the from-scratch model (ModelC). Although both models produced accurate automated segmentations, no significant difference was found between the two training strategies (paired Wilcoxon test, n = 24, p = 0.243). Despite a dataset limited to 20 individuals and the absence of one method clearly outperforming the other, the models remain usable across the three species studied (Gryllus bimaculatus, Loxoblemmus equestris, L. doenitzi), including in the presence of pronounced sexual dimorphism. It reduces a 20 hour segmentation task to under a minute.