A Deep Learning approach for automated Cytoplasm and Nuclei cervical segmentation. (March 2023)

• Record Type: Journal Article
• Authors: Del Moral-Argumedo, Marco J.; Ochoa-Zezzati, Carlos A.; Posada-Gómez, Rubén; Aguilar-Lasserre, Alberto A.
• Is Part Of: Biomedical signal processing and control. Volume 81(2023)
• Publication Date: 2023-03
• Subjects: Deep learning -- Cytological image segmentation -- Pap smear -- Feature pyramid network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104483
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Cervical cancer is a severe disease that affects many women in developing countries. Increasing screening capabilities is the best tool to reduce cancer incidence and save lives. Segmentation is an essential task in screening because it can lead to a better understanding of the morphological characteristics of cells. This paper presents a method for multi-class cell segmentation into Nuclei and Cytoplasm regions. Using the Herlev dataset for evaluation, this work achieved good performance by using state-of-the-art classification architectures such as EfficientNet combined with Feature Pyramid Networks to complete the segmentation task. Performance metrics for both classes show that the approach is robust enough to complete the task. The model achieved a 0.91 F1 score, 0.85 IoU, 0.91 Precision, 0.92 Recall, and 0.96 Specificity as class average with a very low standard deviation, validated with a 5-fold cross-validation. The proposal can help experts correctly asset cervical cell lesions and provide better healthcare for the patients. Highlights: A novel architecture for cervical cell segmentation is proposed. The model can segment cervical cells into both of their components: cytoplasm and nuclei. Use of state-of-the-art Deep Learning techniques such as EfficentNet and Noisy-Student. Balanced model complexity, training, and inference times. Simple inference process to increase adoption in a practical setting.