Refining cell classification for cervical cancer detection using an improved high dimensional feature fusion approach

Early and accurate classification of cervical cell images is essential for timely detection and prevention of cervical cancer. Traditional cytological analysis methods, such as manual interpretation of Pap smears, are often labor-intensive and susceptible to human error. Recent advances in deep learning offer promising solutions, yet many existing models lack generalization across datasets and struggle with multi-class classification challenges.

To address these limitations, this work introduces a Modified High-Dimensional Feature Fusion (HDFF) framework. The proposed method integrates normalized feature vectors extracted from seven diverse pre-trained CNN architectures-VGG16, VGG19, ResNet50, XceptionNet, InceptionV3, DenseNet121, and a Lightweight Feature extractor. These features are concatenated to form a unified representation, which is then processed by a fully connected classifier with dropout and batch normalization to enhance generalization and reduce redundancy.

The model is evaluated on four benchmark datasets: Herlev, SIPaKMeD, Mendeley LBC, and Malhari. It achieves an accuracy of up to 99.85 % in binary classification tasks and maintains high precision, recall, F1-score, specificity, and AUC in more complex multi-class settings. On the Herlev dataset, for example, it attains a precision of 0.995, recall of 0.987, and F1-score of 0.985. Compared to existing approaches, the Modified HDFF demonstrates lower misclassification rates and stable performance across class imbalances and dataset variations.

The results confirm the robustness and adaptability of the Modified HDFF framework, making it a reliable candidate for real-world cervical cancer screening. Its ability to generalize across datasets underscores its clinical relevance and diagnostic value.