Cervical cell classification plays a critical role in the early diagnosis of cervical cancer. While deep learning has achieved remarkable progress in medical imaging, most existing models rely on single-stream feature extraction and do not fully exploit the complementary strengths of different architectures. This study proposes a dual-stream classification framework that leverages lightweight and domain-specialized networks, combined with multiple feature fusion strategies. Our pipeline consists of two parallel branches. The first extracts features using a pretrained ShuffleNetV2, followed by classical machine learning classifiers including SVM, Naive Bayes, Random Forest, KNN, and ANN. The second employs a customized version of CervicalNet, where Grouped Convolution layers are modified to maintain group-wise independence across layers and delay feature fusion until the final stage. The extracted features are reduced using PCA before classification. We further investigate multiple fusion strategies—CCA, feature concatenation, PCA-based fusion, and self-attention—to combine the strengths of both branches. All experiments are conducted on the SIPaKMeD dataset using the MONAI framework. Results show that our MONAI-based implementation improves over baseline reproductions and achieves competitive performance compared to the original published results. Notably, feature concatenation followed by PCA yields the best overall performance, with an F1-score of 0.969 using ANN. This study demonstrates that carefully designed dual-stream architectures with customized grouped convolution and optimized fusion strategies can significantly improve cervical cell classification accuracy.