We present a large-scale dataset of 350 histologic samples from seven different canine cutaneous tumors. All samples were obtained through surgical resection due to neoplastic indicators and were selected retrospectively from the biopsy archive of the Institute for Veterinary Pathology of the Freie Universität Berlin according to the state of preservation and presence of characteristic tumor features. Samples were stained with a routine Haematoxylin & Eosin staining and digitized with two Leica linear scanning system at a resolution of 0.2533 um/pixel. Together with the 350 whole slide images, we provide a database consisting of 12,510 polygon annotations for seven tumor classes and six healthy tissue classes (epidermis, dermis, subcutis, bone, cartilage, a joint class of inflammation and necrosis, melanoma, mast cell tumour, squamous cell carcinoma, peripheral nerve sheath tumour, plasmacytoma, trichoblastoma and histiocytoma).

Due to the large size of the dataset and the extensive annotations, it provides a good basis for segmentation and classification algorithms based on supervised learning. Previous work [1-4] has shown, that due to various homologies between the species, canine cutaneous tissue can serve as a model for human samples.  Prouteau et al. have published an extensive comparison of the two species especially for cutaneous tumors and include homologies between canine and human oncology regarding "clinical and histological appearance, biological behaviour, tumor genetics, molecular pathways and targets, and response to therapies" [1]. Ranieri et al. highlight that pet dogs and humans share many environmental risk factors and show the highest risk for cancer development at similar points of time respective to their life spans [2]. Both, Ranieri et al. and Pinho et al. highlight the potential of using insights from experiments on canine samples for developing human cancer treatments [2,3]. From a technical perspective, Aubreville et al. have shown that canine samples can be used to aid human cancer research through the use of transfer learning methods [4].

The polygon annotations were generated using the open source software SlideRunner (https://github.com/DeepPathology/SlideRunner). Within SlideRunner, users can view whole slide images (WSIs) and zoom through their magnification levels. Using multiple clicks or click-and-drag, the pathologist annotated polygons for 13 classes (epidermis, dermis, subcutis, bone, cartilage, a joint class of inflammation and necrosis, melanoma, mast cell tumour, squamous cell carcinoma, peripheral nerve sheath tumour, plasmacytoma, trichoblastoma and histiocytoma) on 287 WSIs. The remaining WSIs were annotated by three medical students who were supervised by the leading pathologist who later reviewed these annotations for correctness and completeness. 

Potential users of the dataset can load the SQLite database into their custom installation of SlideRunner and adapt or extend the database with custom annotations. Furthermore, we converted the annotations to the COCO format, which is commonly used by computer scientists for training neural networks.

With 350 subjects, the dataset is larger than all currently listed TCIA datasets on cutaneous tumors. Furthermore, its pixel-level annotations can be used for supervised segmentation algorithms as opposed to existing datasets which only provide clinical data on slide-level.