Optimal transport-based synthetic histology generation using cytomorphological Wasserstein barycenters for rare cervical malignancy classification

Despite its potential, the automated cervical cytology analysis is severely limited by the lack of rare malignant cell samples and the class imbalance that is frequently associated with them, resulting in biased predictions and limited generalisation capabilities to clinically relevant abnormal categories. The existing augmentation based methods and generative methods, such as SMOTE, GANs, and variational autoencoders, have major limitations including unrealistic cellular morphologies, mode collapse, feature blur, and lack of biological or statistical assurance, which makes them less reliable for medical decision support. To overcome these problems, this work proposes a novel Optimal Transport based synthetic histology generation framework, Wasserstein Cellular Barycenter Optimal Transport (WCB-OT), that consists of entropic Sinkhorn regularized optimal transport, cytomorphological transport costs, Wasserstein barycenter synthesis and dynamical optimal transport interpolation, to create biologically consistent rare cervical malignancy samples. In contrast to GAN-based, VAE-based and diffusion-based augmentation methods, our proposed method contains no adversarial learning, latent reconstruction, or iterative denoising; instead, it synthesizes a representation by explicitly modeling a distribution transport based on the morphology, and Wasserstein barycentric interpolation. Experimental evaluation on the SIPaKMeD cervical cytology dataset demonstrated that WCB-OT achieved an F1-score of 0.94, AUC-ROC of 0.98, and FID of 1.05, compared with 0.78, 0.91, and 2.67, respectively, for StyleGAN2, the strongest GAN-based baseline, and 0.74, 0.86, and 4.26, respectively, for the SMOTE interpolation-based baseline. The improvements were statistically significant across five independent experimental runs ( p < 0.001). Moreover, the proposed framework offers interpretable transport mappings and biologically meaningful intermediate (cell to cell) transformations that facilitate explainable and clinically reliable synthesis of data. The results indicate that the proposed WCB-OT framework provides a biologically sound, clinically feasible, and theoretically sound approach to enhancing the reliability and diagnostic power of AI-based cervical cancer screening systems for rare cervical malignancies.