Novel prediction model on OSCC histopathological images via deep transfer learning combined with Grad-CAM interpretation. (May 2023)

• Record Type: Journal Article
• Authors: Afify, Heba M.; Mohammed, Kamel K.; Ella Hassanien, Aboul
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: Prediction model -- OSCC histopathological images -- convolutional neural network (CNN) -- gradient-class activation mapping (Grad-CAM) -- ResNet-101; 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.2023.104704
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: In practice, pathologists have to review numerous images of OSCC histopathological, which takes time and is laborious. Although the application of CNN techniques has been successfully applied to other types of cancer, the use of histopathological images of oral cancer has yielded little work. This paper adopted a novel application on the OSCC histopathological image to distinguish between cancerous and normal oral cells in terms of classification and validation. For experimental results, ResNet-101, with the highest accuracy of 100% at 100x magnification, and EfficientNet-b0, with the highest accuracy of 95.65% at 400x magnification, indicate a high performance in predicting oral cancer compared to other modern trained models. The findings gained from this propounded model serve as the crucial leadership of the clinical community in the early accurate detection of oral cancer. Abstract: This paper proposes a novel model using deep transfer learning to predict oral squamous cell carcinoma (OSCC) histopathological images with gradient-class activation mapping (Grad-CAM) to locate the lesion area in the images. The proposed model utilizes a recent public database of 1224 oral histopathological images of normal and OSCC cells at 100x and 400x magnifications. It is inconsequential to base a decision regarding OSCC prognosis on human evaluation, so an accurate decision needs a deep transfer learning strategy that performs better than other comparisons. It may be possible to solve the problem of predicting oral tumors by exploiting the public source database with a proposed model to reach two main targets. The first target contains ten well-classified algorithms based on a deep learning convolutional neural network (CNN) to build a prediction model for OSCC histopathological images to improve classification accuracy between malignant and normal images. After estimating the performance of the models, the obtained results are compared and the best-performing model is selected. The second target contains a Grad-CAM validation process to locate the lesion area in an OSCC image according to the best model. This validation phase can directly impact the robustness of the entire prediction model for OSCC histopathological images. For experimental results, ResNet-101, with the highest accuracy of 100% at 100x magnification, and EfficientNet-b0, with the highest accuracy of 95.65% at 400x magnification, indicate a high performance in predicting oral cancer compared to other modern trained models. The contribution of this work is based on a combination of CNN and Grad-CAM to handle OSCC histopathological images for the purpose of both classification and validation. The findings gained from this propounded model are severe as the crucial leadership of the clinical community in the early accurate detection of oral cancer.