Transformer-based Multiclass Classification of Cervicogram Images for Improved Cervical Cancer Detection

Cervical cancer is a primary health concern that affects women worldwide. Additionally, it is a significant factor in mortality from cancer. On the other hand, when detected early and thoroughly controlled, it is one of the most efficiently treated kinds of cancer. The Human Papilloma Virus (HPV) causes this type of cancer, which promotes abnormal cell proliferation in the cervix region. The Papanicolaou (Pap) smear analysis is the most commonly used method for detecting cancer early. In developed countries, it has had a significant impact on reducing deaths related to cancer. However, underdeveloped countries still struggle to provide affordable solutions owing to a lack of easily accessible medical facilities, and the detection process is time-consuming, irritating, and expensive. Cervicography, on the other hand, may be utilized for mass screening since it is easily accessible, painless, and takes less time. This study proposes an automated cervical cancer diagnosis approach based on Deep Learning techniques in medical image processing to enable mass screening and make cervical cancer detection accessible and affordable to the public. Vision Transformer (ViT) with pre-trained checkpoint from DINO retrieves deep-learned features from cervicogram images delivered into the model by leveraging transfer learning and fine-tuning. This classification project uses the Kaggle Intel and MobileODT dataset. The Kaggle dataset’s three-class classification accuracy rates are 78.7 % (train), 54 % (valid), and 57.9 % (test), respectively.