Machine learning and automation methods for the segmentation, classification and quantification of testicular tissue sections.

Graphical Abstract: Abstract: Analysis of immunofluorescent images can be time-consuming and observer-dependant. We aimed to develop a standardised method for analysing immunofluorescent images of human and mouse testicular tissues using an existing software package. We used QuPath to train and apply an artificial neural network with multilayer perception (ANN-MLP) on images. We were able to automate the segmentation of cells in regions of interest (ROIs) using both StarDist (a convolution neural network) and Watershed image transformation. Segmented cells were classified using QuPath object classification system, which was successfully applied to a range of mouse and human interstitial, Sertoli and spermatogonial germ cell markers. We found that manual counting and classification of cells decoupled the relationship between tubular area and a number of SOX9+ (r 2 = 0.26, P = 0.35) and MAGE-A+ (r 2 = 0.26, P = 0.35) cells. However, automating the segmentation of ROIs and cell classification with simple macros yielded results that maintained the correlation between tubular area and number of SOX9+ (r 2 = 0.56, P = 0.03) and MAGE-A+ (r 2 = 0.93, P = 0.002) cells. In addition, we were able to export data into R/RStudio allowing for the analysis of classification-specific parameters such as mitotic index and cellular organisation in different regions of the tubule. Importantly, the time taken per image using the automated method was significantly faster (6,247 vs three seconds; P < 0.001) at segmenting tubules and quantifying cells than previous manual annotation methods. We propose the use of this method for analysis and cell quantification in testicular tissues. Lay summary: Machine learning (ML) is a type of algorithm that forms part of artificial intelligence (AI). ML is able to learn, detect and predict sequences and structures such as words in sentences or objects in images. ML combined with the ability of modern computers to process large quantities of data and perform repetitive tasks in an automated way makes it an attractive research tool. In particular, the analysis of images taken from a microscope of patient or experimental samples is one area in which ML can excel. We found that open-source software containing ML could be trained on as few as six images. Once trained, the machine learning algorithm could analyse an image in approximately one minute. The same image would take a skilled researcher nearly two hours to analyse. In addition to speed, ML was able to do this more accurately and consistently as well as being automated by a simple piece of code.