Classification of histopathological gastric images using a new method

Abstract

The number of patients who die from stomach cancer is still very excessive. Early diagnosis of cancer patients is necessary to reduce the death rate. In this case, a molecular structure classification system can help carry out early diagnosis of cancer. The existing system has some problems. These problems are high cost, low accuracy rate and waste of time. To tackle this problem, this article presents a new method. This method consists of a new size reduction algorithm called as extending extract histogram of oriented gradients (EEHOG). A multimodality size reduction method is obtained by combining the new EEHOG method and other dimension reduction methods (ODRM). Other dimension reduction methods are multidimensional scaling (MDS), Sammon mapping, Isomap, local linear embedding (LLE), Laplacian eigenmaps, stochastic neighbor embedding (SNE) and t-distributed stochastic neighbor embedding (t-SNE). As well, extract histogram of oriented gradients (EHOG) features have been calculated in this article. Then, EEHOG + MDS, EEHOG + Sammon mapping, EEHOG + Isomap, EEHOG + LLE, EEHOG + Laplacian Eigenmaps, EEHOG + SNE and EEHOG + t-SNE methods have been used for the dimensional reduction of the EHOG features. Thus, the high dimension of these features has been reduced to lower dimensions with a multimodality size reduction method. These new nominal feature sizes have been given to multilayer perceptron neural networks (MLP) and random forest (RF) classifier entries for classification of the histopathological stomach images. The accuracy results established by using the EEHOG + ODRM size reduction method are higher than the accuracy results obtained by using only ODRM. The accuracy result obtained with the MLP classifier is found to be higher from the accuracy results obtained with the RF classifier, when the performance of the MLP classifier is compared with the performance of the RF classifier. The stomach cancer images used this article were obtained from Firat University Medical Faculty Pathology Department. It has been found that the highest accuracy results are obtained as 98.14% with EHOG_EEHOG + LLE_MLP method. The average of the accuracy results obtained with the MLP classifier has been found as 89.57%. However, the average of the accuracy results obtained with the RF classifier has been found as 84.49%. These results have been compared with previous studies. It has been proven that more high accuracy results than single solutions of the multimodality solutions are given.