Development of the Modified Method Based on Convolutional Neural Network of Cancer Cell Nucleus Image Localization

In Taiwan, health insurance payments for cancer treatment are determined based on the patient's recovery. After medical personnel obtains a patient's cell examination results, they can check the decrease in the number or atrophy of cancer cells in the patient through methods such as flow cytometry. Medical personnel generally use fluorescence microscopes to view and count the number of nuclei. However, this method is time-consuming, has a high error rate, and the inspection results are highly inconsistent. Previous studies used convolutional neural networks for cell nuclei localization, automatic counting, and micronucleus analysis to solve the aforementioned problems. However, convolutional neural networks (YOLOV4) are to mis-positions of small-scale dual-nucleus cell images. In this study, the image geometric analysis algorithm is proposed to solve this problem. Using this method, YOLOV4 is used to perform 20X optical magnification for small-scale cell nuclei image localization, and the proposed algorithm was modified to improve the accuracy of cell nuclei localization. To demonstrate small-scale nucleus image localization problems and verify the efficacy of the proposed modified method, the results of the localization of small-scale nucleus image of the YOLO and Faster R-CNN algorithms were compared. The proposed method is shown to correct cell nucleus localization errors. This paper describes the proposed method structure and process in the following sections.