HiAt-Net: A Novel Three-Layered Encoder-Decoder with Hierarchical Attention for Classifying Cervical Cancer with Cytology Images

Cervical cancer (CeCA) is a major worldwide health concern, particularly in low- and middle-income countries where screening and treatment facilities are not widely available. Cervical cells, which connect the lower portion of the uterus to the vagina, are where CeCA, a particular type of cancer, originates. Sexually transmitted infections such as certain strains of the Human Papillomavirus (HPV) are the main cause of cervical cancer. But not every HPV infection results in CeCA; other risk factors include smoking, having several sexual partners, and having a weaker immune system.. Additionally, the classification of CeCA using Deep Learning (DL) can pose challenges related to computational resources, model complexity, and robustness compared to state-of-the-art methodsOur solution, HiAt-Net, tackles these challenges by introducing a 3-layer encoder-decoder approach that incorporates a hierarchical attention mechanism. This method is specifically designed to extract features effectively from complex input data, such as the Herlev dataset and SIPaKMeD dataset. By integrating the attention mechanism at multiple levels, our model can capture hierarchical dependencies and selectively focus on relevant features, resulting in enhanced performance in classification tasks. We conducted a thorough comparison and analysis of our proposed method against existing approaches, evaluating performance metrics such as accuracy, precision, recall, and F1-Score.