<p>Histopathology is an area of medicine that studies diseases through the microscopic observation of tissues and cells. In recent decades, digital pathology has revolutionized histopathology by enabling automatic analysis of tissue images through digitalization of histological samples. Automatic analysis of these images corresponds to computational methodologies often based on supervised machine learning that extract information associated with the tissue. These methodologies have proven to be a valuable support tool for pathologists, improving their accuracy and efficiency in histological analysis and diagnosis. Unfortunately, labeling data in histopathology is tedious, time-consuming, and requires the knowledge of expert pathologists. Often, all the variability in histological images is not properly captured by the labeled data used to train the supervised models. In this scenario, the alternative of unsupervised learning arises. Our work explores the generalizability and robustness of representations obtained in unsupervised segmentation tasks for histopathological images. Particularly, we explore the effect of two plug-in components that can easily be adapted to any unsupervised segmentation model: The Mumford–Shah loss and the median pooling layer. We used 165 images from the Gland Segmentation in Colon Histology Images Challenge to train and test the unsupervised models. Our results show that both mentioned techniques can improve the quality of the segmentations generated by the unsupervised models, resulting in an increase in 0.08 (0.07) SSIM for the median pooling layer (Mumford–Shah loss) in the evaluation dataset.</p>