Analysis of testis histology is fundamental to the study of male fertility, but it is a slow task with a high skill threshold. Here, we describe new neural network models for the automated classification of cell types and tubule stages from whole-slide brightfield images of mouse testis. The cell type classifier recognizes 14 cell types, including multiple steps of meiosis I prophase, with an external validation accuracy of 96%. The tubule stage classifier distinguishes all 12 canonical tubule stages with external validation accuracy of 63%, which increases to 96% when allowing for ±1 stage tolerance. We addressed generalizability of SATINN, through extensive training diversification and testing on external (non-training population) wildtype and mutant datasets. This allowed us to use SATINN to successfully process data generated in multiple laboratories. We used SATINN to analyze testis images from 8 different mutant lines, generated from 3 different labs with a range of tissue processing protocols. Finally, we show that it is possible to use SATINN output to cluster histology images in latent space, which, when applied to the 8 mutant lines, reveals known relationships in their pathology. This work represents significant progress towards a tool for robust, automated testis histopathology that can be used by multiple labs.