Abstract:

The identification of leucocyte, also named white blood cells, types in histological blood tissue images is significant because it enables an opportunity for the diagnosis of various hematological diseases. In this study, for the diagnosis of lymphoma cancer, a hematologic disorder, we presented automatic detection and classification model using a deep learning approach. Faster R-CNN, which is a kind of region-based Convolutional Neural Network (CNN) model, achieves satisfactory performance on object detection and classification problems. To dispose of the feature extraction process in image-based applications, we offer a ResNet50 modified Faster R-CNN model for the detection and classification of leucocyte types which are lymphocyte, monocyte, basophil, eosinophil, and neutrophil in histological blood tissue images. In parallel with this purpose, a novel Faster R-CNN object detection model was designed by modifying ResNet50 model and the locations of leucocytes in the image were determined and classified. The efficiency of the proposed model was tested on a novel histological dataset including blood tissue images. The number of lymphocytes in the blood tissue is used as an evaluation criterion in the diagnosis of lymphoma cancer. Therefore, this study sets an example for clinical studies. According to the proposed model, firstly, the blood tissue images are normalized, and the implicit features are extracted by using the trainable convolution kernel. Then, for the reduction of the extracted implicit features, the maximum pooling is applied. After that, Region Proposal Networks (RPNs) are used to generate high-quality region proposals, which are used by Faster R-CNN for detection. Finally, the softmax classifier and regression layer are carried out to categorize the leucocyte types and estimate the boundary boxes of the test samples, respectively. Experimental results show the successful performance and the generalization capability of novel Faster R-CNN for the detection and classification of leucocyte types. This model demonstrates the potential to be deployed as a diagnostic tool for clinical studies because the method has been tested on a real-world histological data set.

Keywords: