JoWUA

Abstract

Cervical cancer is the second most common cancer affecting women worldwide and results from persistent infections of high-risk HPVs by virtue of their oncogenic effects of the E6 and E7 proteins. In this paper, a new technique for diagnosing cervical cancer based on biomarkers and genes such as KRT17 and CRISP2, and a new IKRL-DBN framework that improves upon conventional models by mitigating the vanishing gradient problem through weighted layer-wise information reinforcement, are proposed. The model improves upon existing models by combining weighted layer-wise information reinforcement with gradient descent, making it more robust and efficient than other frameworks. In addition to the statistical analysis, it has been shown that IKRL-DBN achieved a remarkably high accuracy rate of 95.0% relative to other baseline models such as DBN, CNN, SVM, and Random Forest. Notably, the IKRL-DBN achieved a relatively high recall score, 10.5% higher than that of SVM, which is highly significant for reducing false negatives. From the ablation tests, it was confirmed that domain-specific features play an essential role in prediction since removing them caused the accuracy to decrease by 5.0%. The test also verified that the weighted reinforcement component plays an important role in ensuring the stability and consistency of the learning of the model. Moreover, the statistics for dimensional reduction using PCA and t-SNE showed significant improvements in class separability when using biomarkers and problem-specific weights and relations, with 76.4% to 91.7% improvements. In future research, the model's effectiveness should be verified using more extensive datasets and incorporating other types of data and methodologies.