Classification of Different Tissue Types in Histopathological Images with Conventional and Deep Learning Methods

(I)n recent years, with the increase in computer support, image processing and machine learning methods have become effective on these high resolution histopathological images. The aim of this study is to classify histopathological images with high accuracy to help diagnosis process by reducing the workload of specialists. By using two different data sets in the experimental studies, the images in one of them were classified by conventional feature extractions based method and the images in the other by deep learning methods. In the feature extraction based approach, when pixel densities and image histograms were added to the features obtained with Haralick textural descriptor, the highest classification accuracy was obtained with random forest classifier as 87.3%. In order to overcome the many challenges of feature extraction -based approaches, deep learning methods become as important alternatives. DenseNet121, DenseNet169 and SqueezeNet 1.0 models obtained 88.75%, 92.5% and 88.75% accuracies respectively, by retraining the models trained with the ImageNet dataset provided by PyTorch with the BACH ICIAR dataset and testing them with 20% of the images. With forming an ensemble model, the class which is the majority of the estimation results of the models is counted as the final prediction and 97.5% accuracy is obtained.