Cervical cell's nucleus segmentation through an improved UNet architecture.

Area of Science:

• Medical image analysis
• Computational pathology
• Deep learning for medical imaging

Background:

• Accurate nucleus segmentation is crucial for computer-aided diagnosis (CAD) in cervical cytology.
• Challenges include clumped cells, color variation, noise, and fuzzy boundaries in cervical cell images.
• Deep learning offers a promising approach for overcoming these segmentation difficulties.

Purpose of the Study:

• To propose C-UNet, a novel deep learning model for segmenting cervical nuclei.
• To address challenges posed by overlapped, fuzzy, and blurred cervical cell smear images.
• To evaluate the effectiveness of C-UNet on complex cervical cell datasets.

Main Methods:

• The C-UNet model integrates a bi-directional feature pyramid network (BiFPN) and a wide context unit in its encoder.
• The decoder features two interconnected decoders for mutual optimization and feature integration.
• Data augmentation techniques were utilized to enhance model training.

Main Results:

• C-UNet demonstrated superior performance compared to existing models like CGAN, DeepLabv3, Mask-RCNN, and FCN.
• Achieved high accuracy metrics on complex cervical cell datasets and ISBI-2014/2015 datasets.
• Key results include 93% object-level accuracy, 92.56% pixel-level accuracy, 95.32% object-level recall, 92.27% pixel-level recall, 93.12% Dice coefficient, and 94.96% F1-score.

Conclusions:

• The proposed C-UNet model effectively segments cervical nuclei in challenging images.
• C-UNet offers a significant advancement for automated nucleus segmentation in cervical cytology.
• The model's performance highlights its potential for improving computer-aided diagnosis systems.