Abstract: To support the challenging task of early epithelial cancer diagnosis from in vivo endomicroscopy, we propose a content-based video retrieval method that uses an expert-annotated database. Motivated by the recent successes of non-medical content-based image retrieval, we first adjust the standard Bag-of-Visual-Words method to handle single endomicroscopic images. A local dense multi-scale description is proposed to keep the proper level of invariance, in our case to translations, in-plane rotations and affine transformations of the intensities. Since single images may have an insufficient field-of-view to make a robust diagnosis, we introduce a video-mosaicing technique that provides large field-of-view mosaic images. To remove outliers, retrieval is followed by a geometrical approach that captures a statistical description of the spatial relationships between the local features. Building on image retrieval, we then focus on efficient video retrieval. Our approach avoids the time-consuming parts of the video-mosaicing by relying on coarse registration results only to account for spatial overlap between images taken at different times. To evaluate the retrieval, we perform a simple nearest neighbors classification with leave-one-patient-out cross-validation. From the results of binary and multi-class classification, we show that our approach outperforms, with statistical significance, several state-of-the art methods. We obtain a binary classification accuracy of 94.2%, which is quite close to clinical expectations.