Conceptual Design Of A Hybrid Deep Learning Model For Classification Of Cervical Cancer Acetic Acid Images

Abstract

Automated image-based cervical cancer detection plays a vital role in diagnosing cervical cancer, particularly through the use of digital cervical images obtained via visual inspection with acetic acid (VIA). Many algorithms have been developed to classify these images by extracting mathematical features. Artificial intelligence (AI) has significantly advanced healthcare by improving disease detection, diagnosis, and prediction of health outcomes. While various cervical cancer screening methods have evolved, VIA remains one of the most feasible options in low-resource settings. However, its effectiveness relies heavily on the examiner’s experience, which can be limited due to a shortage of qualified healthcare professionals. This study evaluates the performance of AI image processing techniques for detecting cervical cancer using VIA images. The research compares four traditional machine learning techniques and six deep learning techniques in classifying cervical cancer images, where each model was trained on four randomly selected batches of images (300, 700, 1000, and 1678 images) to assess model performance with an increasing number of training images. The VGG19 model achieved a consistent accuracy of 81% across all training sizes. The Vision Transformer (ViT) model, on the other hand, showed a performance improvement from 57% accuracy with 300 images to 77% accuracy with 1678 images. The hybrid model, combining VGG19 and ViT, demonstrated superior performance with an accuracy of 86.65%, an AUC of 0.85, a sensitivity of 0.832, and a specificity of 0.8485. This study demonstrates that the hybrid model outperforms individual models, offering a promising solution for cervical cancer detection in low-resource environments.