COMPUTER AIDED DIAGNOSTIC SYSTEM FOR WHOLE SLIDE IMAGE OF LIQUID BASED CERVICAL CYTOLOGY SAMPLE CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORK

Abstract

Cervical cancer screening with Papanicolaou test and liquid based cytology relies on the expertise of the pathologist. Liquid based cytology is proven to be more efficient than conventional Papanicolaou test when it comes to sample preparation and possibility of conducting several tests on the same sample. However, specificity and sensitivity of the test are in the range of the Papanicolaou test accuracy metrics, with false negative results still being the main drawback of these manually performed tests. Advances in technology and availability of digital data have enabled succesfull application of machine learning models in diagnostics. Images of cervical cells are now used as input to different deep learning models currently tested in studies concerning computer aided diagnostic systems. This study explores different architectures of convolutional neural network for cervical cancer detection based on Optomagnetic imaging spectroscopy and liquid based cytology samples. The proposed VGG16 based model achieved 93.3% sensitivity and 67.8% specificity in the binary classification problem. Results highlight the need for more balanced dataset in order for suggested deep model to achieve better performance.