Introduction: Microdissection Testicular Sperm Extraction (Micro-TESE) is a surgical procedure that relies on the identification and sampling of wider, more opaque seminiferous tubules to improve sperm yield. However, these discriminators are of limited utility in some cases of azoospermia due to spermatogenic dysfunction (ASD), such as maturation arrest, where tubules can have uniform diameters with highly variable sperm counts. This study evaluates a convolutional neural network (CNN) trained on tissue sperm counts to identify locations of highest sperm yield in a testis in real time.
Methods: Spermatogenic dysfunction was induced in Sprague Dawley rats using one of two methods: testosterone implants (n=12, control n=12) or injections of busulfan (n=6, control n=2), an alkylating, antineoplastic agent. Subjects underwent Micro-TESE and tissues were quantitatively analyzed using microgrid cell counters under phase contrast microscopy. A CNN model was trained on images taken during biopsy and the corresponding sample sperm counts. The model was tasked to predict locations of sperm with > 10 sperm/grid. The neural network model was evaluated with a 111 image dataset. Predictions with <90% confidence were discarded.
Results: The CNN identified sperm dense regions with a sensitivity of 87.5%. Precision was 80%, the F1 score was 83.3%, accuracy was 75%, and false negative rate was 12.5%.
Conclusions: This model advances progress towards an automated targeting system for sperm retrieval by training a neural network using observed sperm count instead of expert opinion. Using a deep neural network backbone, this CNN model incorporates shape and spatial orientation in addition to tubule width and opacity. Our findings suggest the use of an animal ASD model that closely resembles humans is a viable method to train neural networks, and that the new physical discriminators generated by this CNN may apply to human Micro-TESE. Source of
