TOWARDS PRECISION ONCOLOGY: ENHANCED MULTISCALE DEEP FEATURE LEARNING FOR CERVICAL CELL SEGMENTATION

Over 300,000 people die from cervical cancer, which is the fourth most frequent malignancy in women worldwide. The early detection of cervical cancer corelates with significantly improved survival rates and the disease is largely preventable. It is indeed, a rare disease in affluent countries with robust immunization and screening initiatives. Nevertheless, the disease disproportionately impacts women in low- and middle-income nations, who often endure severe and untreatable diseases due to resource scarcity. Our objective is to create a reliable deep learning framework capable of distinguishing between normal and malignant morphologies by accurately delineating the cytoplasm and nucleus of the cells. We employ an innovative U-Net model named YOLOv5++ on a publicly accessible annotated dataset of cervical cytology images. Preliminary findings show the accuracy, IoU and Dice score values of 98.33%, 69.68% and 82.13% respectively which offers enhanced accuracy and efficiency compared to various other advanced U-Net methodologies. Thereby, our study demonstrates the enormous potential of deep learning-enabled segmentation as an effective technique to improve and standardize the cervical cancer screening process, reducing human error and enabling early clinical intervention.