Triagem Automática do Exame de Papanicolaou com base em Atenção Visual

Abstract

Cervical cancer is a severe public health problem, being the fourth malignant neoplasm that most affects women. Researchers have proposed computational techniques to assist the clinician in identifying pre-cancerous lesions in Pap smears. The classical approach consists of an exhaustive search for cellular changes in the image, classifying cells with no considering region features under analysis. Commonly, this methodology has its performance reduced by classification failures into regions with cell overlap, noise, and artifacts with no clinical interest. This thesis presents AGMic, which analyzes images of cervical cells from the screening of regions by an attention mechanism. The proposed model identifies conspicuous areas to human specialists, i.e., salients, and then focuses computation on cell classification of these specific sub-images. The acquisition of cytopathologists’ visual attention occurred through eye-tracking experiments, and the collected data made available as a novel database called CRICVA. This thesis also introduces rSCM to model the nucleus texture and its surroundings, as part of an algorithm that associates the salient regions to potentially abnormal cells. Tests on images from the Brazilian Universal Health System (SUS) revealed that AGMic reduces false positives on image classification and achieves Accuracy (ACC) of 0.955 and Kappa concordance index (K) of 0.919 in three classes (negative, low risk, and high risk) using only the three highest salient regions. These results overperformed the classical approach performance, which obtained ACC: 0.775, K: 0.611 when all nuclei are correctly detected. Since fewer cells are processed, the average runtime for AGMIC was ∼ 5.26 seconds per image, with no GPU, representing a time reduction of ∼ 46.54% for region detection based on nuclei segmentation. Complementary experiments on 4,867 liquidbased cytology images from the public dataset DCCL confirmed the potential of AGMic to select clinical interest regions. These findings highlight the proposed model, enabling it to be used in Papanicolaou screening tools in cytopathological laboratories.