Enhanced cervical cancer diagnosis using MobileNetV2 with transformer-based feature refinement and global attention mechanism

Early detection of cervical cancer is crucial to reducing mortality rates associated with this common malignancy in women. Computer-aided design (CAD) techniques are often used to assist physicians in diagnosing this type of cancer early. This research aimed to accurately diagnose cervical cancer using deep learning techniques applied to cervical MRI scans. This study used the KAUH-CCMD dataset, which includes 1,974 cervical MRI images collected from 500 women at King Abdullah University Hospital (KAUH) in Jordan. The researchers implemented a deep learning model based on MobileNetV2, enhanced with Transformer Blocks to analyze relationships between image segments. A global attention mechanism (GAMBlock) was used to identify the most significant features. Bayesian optimization was applied to tune hyperparameters such as dropout probability and learning rate to enhance model performance. The proposed model achieved an accuracy of 85.86% in diagnosing cervical cancer, outperforming several pre-trained models, including MobileNetV2, VGG16, VGG19, InceptionV3, and Xception. The proposed deep learning method demonstrates high potential for assisting physicians in the early detection of cervical cancer. Its success suggests that similar approaches could be applied in the future to address other medical diagnostic challenges.