EpithNet: Deep Regression for Epithelium Segmentation in Cervical Histology Images.

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Cancer diagnostics

Background:

• Automated pathology offers significant advantages for cervical cancer detection, particularly in resource-limited settings.
• Early detection of premalignant cervical cancer is crucial for effective treatment and improved patient outcomes.

Purpose of the Study:

• To introduce EpithNet, a novel deep learning approach for automated epithelium segmentation in digitized cervical histology images.
• To enhance the accuracy and efficiency of cervical cancer screening through advanced image analysis.

Main Methods:

• EpithNet utilizes three regression networks with varying input block dimensions and network depths.
• The model processes digitized cervical histology images for pixel-wise analysis.
• Intermediate concatenation of convolutional layers (EpithNet-mc) was employed to combine network outputs.

Main Results:

• The EpithNet model demonstrated improved pixel-wise probability estimates by maximizing region-based information.
• The EpithNet-mc variant achieved a 94% Jaccard index, outperforming the benchmark model by 26.4%.
• Evaluation was conducted on 311 digitized histology epithelial images.

Conclusions:

• EpithNet provides superior epithelial segmentation performance compared to existing state-of-the-art methods.
• The developed deep learning technique shows promise for improving automated cervical cancer screening.