Comparison of Deep Learning Models for Body Cavity Fluid Cytology Images Classification

Disease detection is successfully performed in many medical fields using medical images and deep learning architectures. Deep learning architectures used in many fields such as brain images, gastroenterological images and chest x-rays support doctors in the detection of diseases. By using deep learning architectures, diseases can be detected more accurately and faster. With faster detection of diseases, treatment can be started more quickly. Early and accurate diagnosis of cytology images is also very important. After the cytology samples are taken from the patients in general health screenings, they should be interpreted by specialist doctors. The aim of this study is to detect benign and malignant cells from cytology images quickly and accurately. In this study, the Body Cavity Fluid Cytology Images dataset, which includes the data of a total of 21 patients with 14 malignant and 7 benign samples, was used. The dataset contains 693 images of 256x192 size. The images in the dataset are trained using the still popular ResNet50, GoogleNet and AlexNet classification architectures. The classification results obtained were compared with other studies using the same data set. As a result of the test processes, AlexNet architecture achieved 97.26%, GoogleNet architecture 98.12% and ResNet50 architecture 99.13% classification accuracy. The ResNet50 architecture with the best classification result achieved 99.25% specificity, 98.75% sensitivity and 99.13% accuracy. Training and testing in this study showed that the ResNet50 architecture classifies cytology images most successfully.