Abstract

The last decade has shown substantial increasing use of Sperm Chromatin Dispersion tests (SCDt) to evaluate DNA-fragmentation in human sperm. SCDt has been proven advantageous over other techniques. However, the visual evaluation remains a subjective component. The goal of this work was to develop an automated, repeatable and objective image analysis method to evaluate human DNA-fragmentation from SCDt images. SCDt was applied to 12 volunteer male sperm samples imaged by bright-field microscopy. Sperm cell cores (heads) were segmented to extract geometric and texture features. Each sperm cell was manually labeled by experts, in order to build a ground truth (GT) dataset to train a support vector machine for fragmented/non-fragmented DNA classification. The overall predictive performance was assessed using three individual datasets: 100% for training, 10-fold cross-validation, and 70% training/30% testing split. We defined and assessed 11 individual datasets simulating the behavior of the classifier for new, unknown samples. Classification accuracy and incorrectly classified instances show agreement between the proposed method and GT. Results show overall accuracy >90% in the three individual datasets, and a false positive rate <7%. We tested with balanced and imbalanced training datasets (regarding the number of fragmented and non-fragmented sperm cells). Results from the imbalanced dataset show better performance. The predictive performance in the new sample test shows average accuracy >95% and a false positive rate ~2%. Evaluation of the differences between reported DNA-fragmentation percentages against GT show an average/maximum of 4.17%/11.35%, close to the 10% maximum error recommended by the WHO manual for clinical laboratories.