cervical nuclei segmentation through synergic conditional generative adversarial network in cervical smear images
Cervical nuclei segmentation is critical for the early detection and accurate diagnosis of cervical cancer. However, this task is challenging due to the presence of clumped nuclei and variations in texture, shape, and contrast. To address these challenges, we proposed a novel synergic conditional generative adversarial network (SCGAN) for cervical nuclei segmentation. The SCGAN integrates densely connected blocks that progressively extract hierarchical features, a Unified Attention Module (UAM) for selective feature refinement and the Scale-Adaptive Feature Integration and upsampling (SAFIU) module for multi-scale feature integration and upsampling, and a synergic discriminator to enhance adversarial learning. The SAFIU module constructs a multi-scale feature pyramid by progressively upsampling across feature levels, effectively retaining fine spatial details critical for segmenting small nuclei. The Scale-Adaptive Fusion (SAF) block further facilitates feature learning by merging high-level features with low-level spatial cues from the encoder, and then forwarding the fused representation to the corresponding decoder stage. On the adversarial side, the synergic discriminator, consisting of ResNet-50 and EfficientNet-B2, is designed for collaborative learning and accelerates convergence with the help of a synergic block. The integration of an Uncertainty-Aware Attention (UAA) mechanism in the synergic block helps the discriminators concentrate on ambiguous or overlapping regions, thereby providing more informative feedback to the generator. Experiments on multiple cervical nuclei datasets demonstrated that the proposed SCGAN outperformed existing methods in terms of sensitivity, specificity, Dice coefficient, and F1-score. By effectively integrating multi-scale features and leveraging adversarial training, our SCGAN achieves more accurate and more consistent cervical nuclei segmentation, paving the way for improved computer-aided diagnosis systems.