Performance Analysis of Neural Network Architectures for Cervical Precancerous Images Classification Based on Texture Features

Cervical cancer remains a leading cause of cancer-related mortality among women globally, particularly in low- and middle-income countries, necessitating effective diagnostic methods for early detection and treatment. This research focuses on the application of Artificial Neural Networks (ANNs), specifically utilizing the Levenberg-Marquardt (LM) algorithm and comparing its performance with that of Conjugate Gradient with Polak-Ribiere Restart (CGP) algorithms, to classify cervical cell images. RGB color space is utilized for feature extraction, utilizing Haar wavelet transforms to capture both texture and spatial information from the individual color channels. This approach enhance the model's ability to differentiate between High-Grade Squamous Intraepithelial Lesions (HSIL), Low-Grade Squamous Intraepithelial Lesions (LSIL), and normal cervical cells. The LM model achieved optimal performance with a training accuracy of 91.1%, a validation accuracy of 87.6%, and a testing accuracy of 85.9%, demonstrating its robustness in identifying relevant patterns within the data. These results highlight the model's effective learning and generalization capabilities, which are critical for medical diagnostics. The findings emphasize the potential of the LM algorithm in improving cervical cancer classification systems, paving the way for more accurate and reliable diagnostic tools in clinical practice.