EdgeNeXt-SEDP for cervical adenocarcinoma HPV-associated and non-HPV-associated diagnosis and decision support.

AIMS: Adenocarcinoma of the uterine cervix exhibits substantial biological and histological heterogeneity, with subtype-specific differences in prognosis and therapeutic response. Conventional classification-based on histopathology, immunohistochemistry, and molecular testing-remains subjective, labor-intensive, and challenging to standardize. This study introduces EdgeNeXt-SEDP, a lightweight deep-learning framework for automated differentiation of HPV-associated (HPVA) and non-HPV-associated (NHPVA) subtypes from histopathological whole-slide images (WSIs).

MATERIALS AND

METHODS

EdgeNeXt-SEDP integrates three synergistic components: a Squeeze-and-Excitation (SE) module to recalibrate channel-wise feature importance, dual-pooling feature fusion to enrich spatial representation, and progressive stochastic depth decay to enhance generalization. The model was trained and evaluated on 49 WSIs from 21 patients using standardized preprocessing, augmentation, and evaluation protocols. Performance metrics included accuracy, precision, specificity, and macro-averaged F1 score, benchmarked against DilateFormer, RepVIT, and EdgeNeXt architectures.

KEY FINDINGS: EdgeNeXt-SEDP achieved 97.63% accuracy, 97.61% precision, 96.98% specificity, and a 97.58% macro-averaged F1 score, while maintaining computational efficiency with 1.9M parameters and 0.2G FLOPs. Ablation analyses confirmed that each module significantly contributed to performance, with the SE module yielding the largest gains. The proposed model consistently surpassed baseline methods without incurring additional computational cost.

SIGNIFICANCE: By delivering high diagnostic accuracy in an efficient architecture, EdgeNeXt-SEDP offers a scalable and reliable solution for reducing interobserver variability and facilitating timely, individualized management of cervical adenocarcinoma. Its compact design supports integration into diverse clinical and resource-limited settings, advancing the application of AI in digital pathology.