Benchmarking CNN Architectures for Cervical Cell Classification: A Unified Transfer Learning Framework on SIPaKMeD

Abstract

Early detection of cervical cancer remains a major public health challenge, particularly in countries where screening programs are underfunded or inaccessible. Despite the growing number of studies applying convolutional neural networks (CNNs) to analyse cytological images, most use heterogeneous experimental protocols, making it difficult to draw meaningful comparisons across architectures. We bridge this gap by evaluating five pre-trained CNN architectures- ResNet50, EfficientNetB0, DenseNet121, InceptionV3, and MobileNetV2- on the SIPaKMeD cervical cell dataset, within a unique, controlled experimental framework. Each model is run with the same preprocessing pipeline, the same transfer learning policy (80%-layer freezing), and the same training hyperparameters. We evaluated each model in two contexts: a five-class cellular classification task and a clinically driven three-class clustering task, using accuracy, macro-precision, macro-recall, and macro-F1-score. Our best result- where DenseNet121 achieved 96.93% ± 0.38 in the five-class framework and 97.47% ± 0.47 in the three-class framework- outperforms all results obtained using a single architecture previously published on this dataset under similar conditions. The framework we describe provides a repeatable starting point for subsequent optimization work.