Validation of an automated counting procedure for phthalate-induced testicular multinucleated germ cells

Abstract

In utero exposure to certain phthalate esters results in testicular toxicity, characterized at the tissue level by induction of multinucleated germ cells (MNGs) in rat, mouse, and human fetal testis. Phthalate exposures also result in a decrease in testicular testosterone in rats. The anti-androgenic effects of phthalates have been more thoroughly quantified than testicular pathology due to the significant time requirement associated with manual counting of MNGs on histological sections. An automated counting method was developed in ImageJ to quantify MNGs in digital images of hematoxylin-stained rat fetal testis tissue sections. Timed pregnant Sprague Dawley rats were exposed by daily oral gavage from gestation day 17 to 21 with one of eight phthalate test compounds or corn oil vehicle. Both the manual counting method and the automated image analysis method identified di-n-butyl phthalate, butyl benzyl phthalate, dipentyl phthalate, and di-(2-ethylhexyl) phthalate as positive for induction of MNGs. Dimethyl phthalate, diethyl phthalate, the brominated phthalate di-(2-ethylhexyl) tetrabromophthalate, and dioctyl terephthalate were negative. The correlation between automated and manual scoring metrics was high (r = 0.923). Results of MNG analysis were consistent with these compounds' anti-androgenic activities, which were confirmed in an ex vivo testosterone production assay. In conclusion, we have developed a reliable image analysis method that can be used to facilitate dose-response studies for the reproducible induction of MNGs by in utero phthalate exposure.

Automated identification of multinucleated germ cells with U-Net

“…(Barlow and Foster 2003); (Kleymenova et al 2005)). However, accurate dose-response data for MNGs have not been calculated for a wide array of phthalates (Spade et al 2018).…”

“…The primary drivers of the low quantity of MNG count data are the time required for performing the count and the training needed for an expert to be able to confidently identify the MNGs (Spade et al 2018). Additionally, there is error inherent in recognizing an MNG in a thin tissue section, or especially a two-dimensional image of a thin section.…”

“…As a result of these features, different experts are able to reach different conclusions about some fraction of the cells that are identified as MNGs. Spade et al (2018) began the process of addressing these issues through the creation of a semi-automated counting pipeline. Using hematoxylin-stained thin sections of fetal rat testes and a scripted process though the NIH ImageJ software, (Spade et al 2018) identified MNGs based on their image characteristics using computer vision principles.…”

“…The primary image criteria used to identify MNGs was size. Without a formal cell segmentation routine, Spade et al (2018) used image thresholding, down-sampling, blurring, and filling to identify connected image components. These connected components were filtered based on circularity to exclude non-cells, and then they were filtered by size to identify MNGs.…”

“…Here, we propose to improve on the Spade et al (2018) approach, leveraging recent innovations in convolutional neural networks to build a fully automated MNG identification code.…”