Automated detection of tumor regions from oral histological whole slide images using fully convolutional neural networks. (August 2021)

• Record Type: Journal Article
• Authors: dos Santos, Dalí F.D.; de Faria, Paulo R.; Travençolo, Bruno A.N.; do Nascimento, Marcelo Z.
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: Oral tumor segmentation -- Fully convolutional neural networks -- H&E-histological image -- Whole slide images; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102921
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: The first FCN-based method to perform refined segmentation of oral tumors in WSIs. The method was applied in different datasets with WSIs of different tumors. The proposal achieved segmentation accuracy results up to 97.6%. Abstract: The diagnosis of different types of cancer, including oral cavity-derived cancer, is made by a pathologist through complex and time-consuming microscopic analysis of tissue samples. This paper presents a method based on a fully convolutional neural network to localize and perform refined segmentation of oral cavity-derived tumor regions in H&E-stained histological whole slide images. The proposed method uses color features in the HSV color model to identify tissue regions in a pre-processing step to remove background and nonrelevant areas. The identified tissue regions are then transformed into the CIE L*a*b* color model and split into image-patches. The method was applied in a WSI dataset of oral squamous cell carcinoma tissue samples. In addition, for further validation and comparison with other proposals, we also applied the proposed method in a WSI dataset of sentinel lymph nodes with breast cancer metastases. Experimental evaluations were performed using a total of 85, 621 image-patches of size 640 × 640 pixels and the proposed method achieved good results in different cancer-derived datasets with images of different tumors. The results revealed that the proposal is robust and capable to localize and perform refined segmentation, achieving accuracy results up to 97.6%, specificity up to 98.4%, and sensitivity up to 92.9%. The influence of different color spaces and different image-patch sizes in the proposed method also were explored.