AMP-Net: An Adaptive Multi-Branch U-Net with Prior-Guided Fusion for HEp-2 Cell Segmentation

Abstract

Accurate segmentation of HEp-2 cell images is a crucial step in indirect immunofluorescence (IIF) analysis for the diagnosis of autoimmune diseases. To maintain computational efficiency, especially for deep learning models, high segmentation efficiency remains a major challenge.
To address this challenge, we propose AMP-Net, a lightweight adaptive multi-branch U-Net with a prior-guided fusion, which is specially designed for effective HEp-2 cell segmentation.
AMP-Net combines MobileNetV2 encoder, multi-branch architecture, adaptive prior fusion, and dual-head learning strategy for simultaneous mask and border prediction.
AMP-Net improves localization and edge delineation by integrating complementary prior information from Otsu thresholding and distance-transform maps, which improve segmentation quality.
Extensive experiments on the MIVIA HEp-2 dataset with 5-fold cross-validation show that the proposed model achieves a mean accuracy of 97.97%, precision of 94.83%, sensitivity of 96.57%, specificity of 98.46%, Dice coefficient of 95.61%, and IoU of 91.65%.
The best fold has a Dice coefficient of 96.76% and an IoU of 93.74%. Without data augmentation, AMP-Net achieves a Dice coefficient of 0.950 +- 0.016, while with augmentation it achieves 0.956 +- 0.017. Similarly, the IoU rises from 0.905 to 0.916.
The results obtained show that AMP-Net consistently maintains accuracy, robustness, and computational efficiency and thus is suitable for practical biomedical image segmentation applications.
Qualitative evaluation also demonstrated the robustness of AMP-Net with Dice scores of 98%, 97%, and 95% for the best, average, and challenging segmentation cases, respectively, while preserving accurate cellular boundaries and structural consistency.
Besides quantitative evaluation, AMP-Net is proposed as a practical segmentation module for HEp-2 computer-aided diagnosis to achieve accurate cell isolation for further ANA pattern classification and quantitative morphological analysis.