Melanoma Histopathology Dataset with Tissue and Nuclei Annotations

Description:

This dataset is designed for development of deep learning models for segmentation of nuclei and tissue in melanoma H&E stained histopathology. Existing nuclei segmentation models that are trained on non-melanoma specific datasets have low performance due to the ability of melanocytes to mimic other cell types whereas existing melanoma specific models utilize older, sub-optimal techniques. In addition, existing models do not provide tissue annotations necessary for determining the localization of tumor infiltrating lymphocytes whereas this might also hold value for predictive and prognostic tasks. To address this we created a melanoma specific dataset with nuclei and tissue annotations. 

Methodology:

• Sample Collection: Region of interest (ROI) were sampled from H&E stained slides of 100 primary melanomas and 100 metastatic melanomas scanned with a Hamamatsu scanner at 40× magnification (0.23 μm per pixel). From each slide a 40× magnified ROI of 1024×1024 pixels was selected for annotation. In addition a context ROI of 5120×5120 pixels was sampled to provide information about the broader context for the annotation process and if needed to be able to generate a larger amount of annotations. Selection was done by a trained medical expert (M.S.) and subsequently verified by an expert dermatopathologist (W.B.). Manual ROI selection ensured diverse tissue and nuclei types.
• Annotation Process: Nuclei segmentations were generated with HoverNet pretrained on the PanNuke dataset. Manual annotation was performed by M.S.
  using Qupath with the following cell categories: tumor, stroma, vascular endothelium, histiocyte, melanophage, lymphocyte, plasma cell, neutrophil, apoptotic and epithelium and tissue categories: tumor, stroma, epithelium, endothelium, necrosis, white background. Annotation categories were based on earlier datasets. In addition, we chose categories based on possible predictive value. All annotations were checked by an expert dermatopathologist (W.B.). 
• Quality Control: To assess the reliability of the annotations, intra- and inter-observer agreement (by experienced dermatopathologist G.B.) were determined on 10 randomly selected ROIs. The intraobserver overall precision was 89.47% with a recall of 97.71%, and an F1 score of 93.41%. Interobserver precision was 84.01% with a recall of 94.87% and an F1 score of 89.11%. These results are based on the sum of all true positive, false positive and false negatives for the 10 ROIs.