Comparison of CNN Architectures for Pre-Cancerous Cervical Lesion Classification Based on Colposopy Images Using IARC and AnnoCerv Datasets

Cervical cancer represents a significant public health issue affecting women worldwide, and identifying the severity of lesions early on is crucial to selecting the right treatment. This research investigates and compares the effectiveness of various Convolutional Neural Network (CNN) models in classifying colposcopic images according to the severity of cervical lesions. The dataset used was obtained from the International Agency for Research on Cancer (IARC) and AnnoCerv, consisting of 452 colposcopy images categorized into four classes: Normal, CIN 1, CIN 2, and CIN 3. Five CNN architectures were evaluated: MobileNetV2, InceptionV3, Xception, VGG16, and DenseNet121. Experiments were conducted using default hyperparameters: batch size of 32, learning rate of 0.001, and 100 epochs. The results showed that MobileNetV2 achieved the highest accuracy at 67%, followed by DenseNet121 (60%), Xception (60%), InceptionV3 (55%), and VGG16 (42%). Based on these findings, MobileNetV2 is the most optimal model for classifying colposcopy images in this study. However, the study is limited by class imbalance and dataset size, which may affect model generalizability. Future work may explore ensemble learning techniques and larger, more diverse datasets for improved accuracy.