BASG-Net: a background-aware spatial gating network for cervical cell segmentation in pap smear images

Area of Science:

• Medical Imaging
• Computational Pathology
• Biomedical Engineering

Background:

• Accurate cervical cell segmentation is crucial for computer-aided cervical cancer screening.
• Pap smear images present challenges due to background interference from non-cellular components.
• Existing segmentation models often struggle with distinguishing cells from background noise.

Purpose of the Study:

• To propose a novel network, BASG-Net, for improved cervical cell segmentation in Pap smear images.
• To explicitly model and mitigate background interference during the segmentation process.
• To enhance the robustness of segmentation models in complex imaging conditions.

Main Methods:

• Introduction of a Background-Aware Auxiliary branch to estimate non-cell/background probability maps.
• Utilization of a spatial gating signal derived from background estimation to modulate feature fusion.
• Incorporation of meso-scale convolutions to improve the representation of cellular structures and context.

Main Results:

• BASG-Net achieved high performance metrics, including Dice coefficients of 0.7719 (APACS23) and 0.9425 (Cx22).
• IoU values reached 0.6286 (APACS23) and 0.8913 (Cx22), with accuracies of 0.9796 and 0.9735, respectively.
• Qualitative analysis indicated a reduction in background-related false positives in complex smear images.

Conclusions:

• The proposed BASG-Net effectively addresses background interference in cervical cell segmentation.
• Background-aware spatial gating is a promising strategy for improving segmentation accuracy in challenging Pap smear images.
• BASG-Net demonstrates significant potential for enhancing computer-aided cervical cancer screening systems.