Deep Residual Squeeze-and-Excitation Network for Colposcopy Image Classification

Due to a lack of screening facilities, knowledgeable professionals, and public awareness, cervical cancer continues to rank among the most common malignancies among women, especially in developing nations. Expert review is frequently necessary for current screening methods such the Papanicolaou (Pap) test, histopathology, visual inspection with acetic acid (VIA), and the human papillomavirus (HPV) test. In this study, we address the challenge of colposcopy image classification by developing a deep convolutional neural network enhanced with squeeze-and-excitation (SE) blocks and residual connections to improve feature extraction and classification accuracy. The model leverages SE blocks to emphasize critical features and residual connections to ensure efficient gradient flow in a deep architecture. We conducted hyperparameter fine-tuning and applied Principal Component Analysis (PCA) to optimize performance. The model’s initial training and validation accuracy were 96.72% and 96.23%, respectively. Training accuracy increased to 98.12% with a validation accuracy of 98.01% after fine-tuning, proving the efficacy of our strategy. These results show that adding SE blocks and residual connections significantly improves the model’s competency to detect cervical lesions in colposcopy images. According to the results, this design may offer strong support for automated cervical cancer screening, which might help with early diagnosis and detection.