Automated segmentation of tumor regions from oral histological whole slide images using fully convolutional neural networks

Abstract

Segmentation of tumor regions in H&E-stained slides is an important task for pathologists while diagnosing different types of cancer, including oral squamous cell carcinoma (OSCC). Whole slide histological images (WSI) are high-resolution images created from tissue sections that can contain many different types of cells, regions, and some artifacts derived from the image acquisition process. Processing these huge images to discard non-relevant areas while keeping only the clinically relevant regions is a challenging task. This work presents a new approach, based on fully convolutional neural networks (FCN), capable of automatically analyzing a WSI to segment regions of oral cavity tumors to aid and enhance the pathologist's decision making on OSCC diagnoses and prognosis. The proposed methodology uses color characteristics in the HSV color space to identify tissue regions in a pre-processing step. The tissue regions identified in the WSI are transformed to the CIE L*a*b* color space and split into smaller sub-images. An FCN segmentation model, based on the U-Net model and trained using a new data augmentation strategy, is used to locate and segment tumor regions from the sub-images. The evaluation of the proposal was tested using public domain image datasets, and a newly created WSI dataset of H&E-stained OSCC tissues. The results have shown a segmentation F1-score of 97% in the new OSCC dataset created. For the public domain image datasets, including a dataset of WSIs from another type of cancer, the proposed method obtained an average F1-score of 83%, proving to be robust and with the potential to develop a tool to support the diagnosis in cases of OSCC.