Monday, August 8th, 2022


A full text version of this article is available.
To access article obtain online access here or login
Analysis of Semantic Segmentation Approaches for the Morphometric Analysis of Feulgen-Stained Cytological Samples
Authors:  Luiz Antonio Buschetto Macarini, B.Sc., Antonio Carlos Sobieranski, Ph.D., Aldo von Wangenheim, Ph.D. (Dr. rer. nat.), Alexandre Sherlley Casimiro Onofre, Ph.D., Fabiana Botelho de Miranda Onofre, Ph.D., and Marcelo Ricardo Stemmer, Ph.D.
  Objective: The main goal is the early detection of neoplasia patterns through the quantification of DNA in Feulgen-stained sample images through deep learning methods.
Study Design:
Comparison of U-Net and Attention U-Net for image semantic segmentation, employing ResNet18 and ResNet34 as their backbones. The presented approaches for both binary and multi-class segmentation tasks were also verified. The evaluation was conducted at two image resolutions: 600×800 pixels and 1200×1600 pixels, resulting in a total of 16 experiments. Our dataset contains 1,010 images.
Best results were achieved with Attention U-Net employing a ResNet18 as its backbone, with 600× 800 as image resolution. This architecture achieved an intersection over union (IoU) of 0.809. In the multiclass segmentation, the best results were achieved with the U-Net, employing a ResNet18 as its backbone, and an image field resolution of 600×800, resulting in an IoU of 0.638.
Semantic segmentation using convolutional neural networks (CNNs) showed to be a robust approach. Experimental results demonstrate the validity of the use of these networks as a promising solution for the automated segmentation of cell nuclei. These approaches have potential to be employed as a step of a possible pipeline for nuclei segmentation.
Keywords:  cell nucleus, cervical cancer, cytodiagnosis, Feulgen stain, neural networks, uterine cervical neoplasms
  Acrobat Reader 7.0 is recommended to properly view and print the article.
Reader can be downloaded from