Supervised learning for semantic segmentation of human spermatozoa

Image-based diagnosis is able to spot several diseases and clinical conditions faster and more accurately than traditional manual ones, becoming also an alternative in monitoring and predicting patients responses to specific health treatments. In this work, we present a supervised learning approach to segment pixel-wise parts of spermatozoa using a random forest (RF) classifier. The framework created a multi-channel image combining intensity RGB bands with three neighborhood based bands. The last neighborhood based bands were Sobel’s magnitude and orientation and Shannon’s entropy. A RF was trained using labeled pixels provided by expert andrologists, biochemists and specialists in reproductive health. We compared results with a simple model on the RGB only. The whole automatic process (segmentation and classification) achieved an average precision of 98%, recall of 98% and F-Score of 98%. Highest improvement in comparison to the RGB model was shown on the segmentation of the tail. We provided a fully automatic spermatozoa semantic segmentation based on local and non-local information. The results are aimed to develop a CASA (Computer Assisted Sperm Analysis) system that can provide results over the Internet. The experiment was conducted on normalized images of a specific microscope. We are planning to extend the experiment in future work to more realistic conditions including different stainings, microscopes and resolutions.