Cluster—GAT: Mixing Convolutional and Self-Attended Feature Maps Using Graph Attention Networks for Cervical Cell Classification

With recent upsurges in Cervical cell classification it has become crucial to design a ubiquitous framework that is capable of accommodating new classes or feature pool, without undergoing any significant alterations in its training methodology. In this paper, we have explored a novel approach of ensembling weakly (supervised) trained feature extraction models to combine the positive aspects of Convolutional Kernels and Attention based Vision Transformers. We have also shown that Graph Attention Networks can resolve classification discrepancies in WSI based datasets by attenuating outliers and generating intra-class relation aware image-level representations. The problem of scarce training samples and class imbalance has also been resolved by using a blend of Generative Adversarial Networks and digital filters. The proposed framework was evaluated on three benchmark cervical cytology image datasets (SIPaKMeD, Herlev, and CRIC) to exhibit the uniformity of performance across all specimen test cases.