Tensor-based Color Feature Image Extraction for Blood Cell Recognition
Ngoc Tuyen Le, Anh Quynh Vu, Hoan Quoc Bui, Long Tuan Nguyen · 2023
Blood cell recognition plays a huge role in medicine. Many researchers have recently promoted the use of deep learning models to improve the accuracy of blood cell recognition. However, many remain challenges, such as the unavailability of large training data sets and the similarity between normal and malignant cells. To overcome these challenges, this paper proposed a novel method to extract the feature of blood cell images in the preprocessing step of a deep learning-based blood cell recognition system. First, the color blood cell image as a third-order tensor is expressed, and then compute its core tensor using high-order singular value decomposition (HOSVD). Second, a new core tensor is created by fixing the second frontal slice, and the remaining frontal slices are set to all-zero matrices. Finally, the feature of the blood cell was obtained by reconstructing the image with the new core tensor by inverting HOSVD. Experiments performed on two most famous public medical datasets named BCCD and Raabin-WBC using three state-of-art deep learning models, AlexNet, GoogLeNet, and ResNet show that by using the proposed feature image, a very high accuracy recognition rate is obtained, higher the using original images even use only small amount data for training.