Automated microscopy in forensic biology: Validating the use of deep neural networks for the detection of human spermatozoa in Christmas Tree stained slides

<p>The most widely accepted technique for the confirmation of semen presence in forensic biology is the microscopic identification of sperm cells. The process of manual microscopic examination, however, can often be time consuming, particularly where sperm numbers are low or where sperm cells are absent, particularly when in the presence of numerous non-sperm cells. The findings of this study demonstrate that automated microscopy utilising Deep Neural Network (DNN) allows for an improvement of the efficiency of examination times associated with microscopy. Additionally, the use of instrument and slide data as diagnostic tools permits rapid detection of invalid results. Metafer with Sperm Detection, using DNN v3.1.1, successfully identified sperm cells present in both high and low semen concentrations in mixed-cell substrates, often with a higher accuracy than manual microscopic examinations where very low sperm numbers were present (1\u8211?5 sperm cells on an entire slide). Whilst there were instances of false positive and false negative sperm cell classifications observed, the overall sensitivity, specificity and accuracy of automated microscopy using Metafer with Sperm Detection (DNN v3.1.1) was 0.9999, 0.9996 and 0.9999, respectively. The conversion of Metafer sperm counts to an existing sperm grading system allows compatibility with manual microscopy outcomes and allows decisions regarding sample triage and submission for specific DNA extraction protocols in routine casework to be made with confidence.</p>

Digital object identifier (DOI): 10.1016/j.scijus.2025.101370