Evaluating Pre-trained Neural Networks in Deep Learning for Early Detection and Enhanced Screening of Cervical Pathology

Abstract

This document presents a research thesis for the Master in Computer Science (MCC) degree at Tecnologico de Monterrey. Cervical cancer remains a leading cause of mortality among women, particularly in low-resource regions where screening tools such as the Pap smear often fall short in early detection. This research explores the application of deep learning and pre-trained neural networks for the binary classification of cervical pathology, focusing on detecting dysplasia, specifically CIN2 and CIN3, as a potential prevention tool. We im- plemented multiple neural network models, including DenseNet, EfficientNet, MobileNet, and ResNet. The models were evaluated on two distinct datasets: one from the International Agency for Research on Cancer (IARC) and another from the Zambrano Hospital. To as- sess the generalization capacity of these models, we employed a sequential training approach where the first batch was trained with IARC data and tested on a Zambrano Hospital batch, with subsequent tests progressively incorporating prior results. Each experiment was repeated over 10 iterations to calculate confidence intervals for the performance metrics. Our results demonstrate that DenseNet and EfficientNet outperformed other models, achieving superior sensitivity and accuracy compared to conventional Pap smear tests. These findings indicate that deep learning models hold promise as an affordable, effective cervical cancer screening tool in low-resource communities. Future work will focus on augmenting datasets through collaboration with healthcare institutions and exploring generative models such as GANs to improve model robustness and generalization.