Performances of Pre-Trained Models in Classification of Body Cavity Fluid Cytology Images

Abstract

Using deep learning architectures, disease detection from medical images such as brain images, chest X-rays, and cytology images is used in many areas. With the use of deep learning architectures in the faster detection of diseases, the possibility of starting treatment early increases. Benign and malignant cell clusters are formed from body cavity effusions. Diagnosis is made by visual inspection of images of centrifuged body fluid effusions deposits. Bad cells are usually abundant in body cavity fluids. Here, separating malignant cells from the proliferation of mesothelial cells or inflammatory cells is the problematic part. In this case, a classifier can be used to distinguish benign mesothelial or inflammatory cells from malignant carcinoma cells. In this study, using Convolutional Neural Network (CNN) based models, the detection of benign and malignant cells from cytology images consisting of fluids in the body cavity will be an aid to pathologists by detecting early and accurately. The dataset contains 693 images with two classes of 256x192 benign and malignant cells. These images are trained using AlexNet, ResNet50, DarkNet53, GoogleNet, MobileNetv2, and EfficientNetb0 architectures. As a result of the test phase, the highest classification accuracy was obtained in the DarkNet53 architecture with 98.56%.