Predicting Cervical Cancer with Deep Learning: Comparative Analysis of VGGNet, GoogleNet, and DenseNet121 Models
Arshleen Kaur, Rishabh Sharma, Richa Gupta, Ashish Garg · 2024
The proposed study compares the efficacy, degree of accuracy, and computational efficiency of three deep learning algorithms, namely, VGGNet, GoogleNet, and DenseNet121, in terms of their capacity to determine cervical cancer from smear images and thus further the role of artificial intelligence in medical tests. The analysis that we conducted quantitatively was on a dataset of cervical smear images annotated with labels such as normal, pre-cancerous, and cancerous. The data set is preprocessed by conducting image resizing, augmentation, and normalizing the data. For each model, I used TensorFlow and Keras frameworks, which were built with cervical cancer in mind. During the training, it splits the data in the training, validation, and test set and monitors the performance with accuracy, precision, recall, Fl-score, and AUC. A qualitative analysis told apart the models based on these metrics in terms of training time and model complexity. DenseNet121 excelled in all image segmentation tasks with the highest accuracy (93%)., precision (90%)., recall (95%), F1-score (92.5%), and AUC (0.97). GoogleNet, although slightly less accurate, was nevertheless the method employing the shortest training time and the least complexity. In this way, it becomes a viable choice for limited resource settings. VGGNet used a little bit, but it did give important ideas about model depth's influence on medical image quality. It furthermore brought to the issue of accuracy-computational demand tradeoffs of the models.