So far, the qualitative and quantitative analysis of bainite has to be carried out by a metallography specialist and often causes ambiguities coming along with poor reproducibility. Possible reasons for a high variance in the description of bainite originate from different expert opinions given the high variety in the appearance of bainite in micrographs. In particular, the applied cooling regime and corresponding temperature gradients in the material dictate the evolving microstructure and its complexity. In recent years, deep learning showed its potential to provide a robust and fast quantification of image data derived from learning large datasets. In order to unfold the potential of deep learning and facilitate its usage for the material science community, a deeper understanding on the role of data pre-processing is necessary to capture the influence of metallography images (and their complexity) on the learning process. In this study, the open-source detection and segmentation library Detectron2 (https://github.com/facebookresearch/detectron2) was used within a framework to quantify a crucial constituent in bainite - the martensite-austenite (M-A) islands - in electron microscopy images. We provide three bainite data sets with image data representing different cooling regimes and therefore different M-A characteristics. From segmentation results, the ratio of constituent to image size manifests as a crucial parameter during pre-processing affecting the accuracy of subsequently trained models.