Fusing Expert Knowledge and Deep Learning for Accurate Cervical Cancer Diagnosis in Pap Smear Images: A Multiscale U-Net with Fuzzy Automata

Abstract

Ovarian cancer is a severe disease that impacts many women in developing countries.  Increasing screening capacity is the most effective strategy for lowering cancer risk and saving people's lives. Early stages of cervical cancer often lack symptoms, making it the fourth leading cause of mortality among women. Although cancer cells grow slowly in the cervix and can be effectively treated if detected early, detecting it before it rapidly spreads are a major challenge for the medical community. Segmentation is a critical screening step as it enhances our comprehension of cell morphological properties. This study provides a technique to segment multi-class cells into Nucleus and Cytoplasmic areas. Multi-resolution U-Net (MRU-Net) is provided for medical image segmentation to bypass the constraints of U-convolution Net's kernel with a restricted receptive field and undetermined ideal network width. First, additional semantic information is extracted from the images using a series of recurrent convolutions.  Second, to distinguish the characteristics, a convolutional unit with distinct receptive fields is utilized. The effects of network width inconsistency may be mitigated by integrating a convolution layer with a large number of receptive fields. The effectiveness of the research was measured against state-of-the-art methods using the Herlev dataset and classification structures were used to get excellent results. Effectiveness indicators for both groups suggest that the method is reliable enough to complete the task. The approach may enable doctors to identify cervical cell anomalies and provide improved medical care. MRU-Net is evaluated using varied medical image segmentation datasets.