A Comparative Analysis Of Pretrained Convolutional Neural Networks For Colposcopy-Based Cervical Cancer Classification Under Imbalanced And Augmented Data Conditions

Abstract

In medical image processing and diagnosis, a robust and generalized deep learning model is extremely important to produce a precision decision. Deep learning models that can sustain stability and accuracy in the context of restricted data availability are essential for reliable automated diagnosis. This work investigates the performance of various pretrained models when trained on both balanced and imbalanced image dataset. The data set is comprised of three classes, CIN1, CIN2, and CIN3, which are colposcopy-based cervical images.

The prediction rate quality metrics for several models with augmentation range from 40.4% to 99.4% across different classes. The quality metrics of prediction rate for various models without augmentation ranged from 40.6 to 98.6% for various classes. The average classification test accuracy prior to augmentation for Class 1, Class 2 and Class 3 are  0.59,0.645 and 0.4, respectively. The average classification test accuracy with augmentation for Class 1, Class 2, and Class 3 is 0.947, 0.99, and 0.99, respectively. The results demonstrate there is an enhancement in the performance of the models with data augmentation. The comparative study on the models is also evaluated to ensure an accurate classification rate of colposcopy-based cervical cancer images.