cervical nuclei segmentation through synergic conditional generative adversarial network in cervical smear images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Cervical nuclei segmentation is crucial for cervical cancer diagnosis.
• Challenges include clumped nuclei and variations in texture, shape, and contrast.

Purpose of the Study:

• To develop a novel synergic conditional generative adversarial network (SCGAN) for accurate cervical nuclei segmentation.
• To address the limitations of existing methods in handling complex cervical cell images.

Main Methods:

• Proposed SCGAN integrates densely connected blocks for feature extraction, a Unified Attention Module (UAM) for refinement, and Scale-Adaptive Feature Integration and upsampling (SAFIU) for multi-scale processing.
• Employed a synergic discriminator with ResNet-50 and EfficientNet-B2, incorporating an Uncertainty-Aware Attention (UAA) mechanism to focus on ambiguous regions.
• Utilized Scale-Adaptive Fusion (SAF) blocks for merging encoder-decoder features.

Main Results:

• SCGAN demonstrated superior performance over existing methods on multiple cervical nuclei datasets.
• Achieved higher sensitivity, specificity, Dice coefficient, and F1-score.
• The model effectively integrated multi-scale features and adversarial training for precise segmentation.

Conclusions:

• The proposed SCGAN offers more accurate and consistent cervical nuclei segmentation.
• This advancement holds significant potential for improving computer-aided diagnosis systems for cervical cancer.
• The integration of advanced deep learning techniques addresses key segmentation challenges.