CerviSegNet-DistillPlus: An Efficient Knowledge Distillation Model for Enhancing Early Detection of Cervical Cancer Pathology

Abstract

The global challenge of cervical cancer calls for advancements in early detection and diagnosis. Our study introduces CerviSegNet-DistillPlus, a state-of-the-art deep-learning framework that elevates cervical cancer cell detection and segmentation. It leverages the DeepLabV3+ architecture, enhanced with leading-edge knowledge distillation and model pruning to efficiently process diverse data and operate within computational limits typical in clinical settings. This results in a compact yet highly accurate model that excels in computational efficiency. In a comparative analysis, CerviSegNet-DistillPlus achieves top performance, improving accuracy by 0.8%, 1.5%, and 2% over its nearest rivals on the Cx22, Technical University of Denmark/Herlev Hospital Pap Smear Database(DTU/HERLEV), and SIPaKMeD datasets, respectively. On the Cx22 dataset, it attains a sensitivity of 0.9623, specificity of 0.9219, accuracy of 0.94, and a top Dice coefficient of 0.9855. For the DTU/HERLEV dataset, CerviSegNet-DistillPlus demonstrates a sensitivity of 0.9617, specificity of 0.91, accuracy of 0.9365, and a remarkable Dice coefficient of 0.9892. Furthermore, on the SIPaKMeD dataset, it achieves a sensitivity of 0.9369, specificity of 0.899, accuracy of 0.9249, and an outstanding Dice coefficient of 0.9734. The integration of knowledge distillation and test-time augmentation significantly improves segmentation accuracy, while model pruning substantially reduces computational complexity, making it well-suited for efficient deployment in clinical settings. This innovative integration of advanced techniques achieves high accuracy and efficiency for cervical cancer cell detection. CerviSegNet-DistillPlus stands as a powerful, efficient, and accessible tool for early cervical cancer diagnosis, offering the potential to improve patient outcomes and make a significant contribution to the global fight against cervical cancer.

Keywords

• Cervical cancer
• deep learning
• knowledge distillation
• model pruning
• computational efficiency
• swin-transformer