Image Processing Methods for Automatic in-vitro Morphology Analysis

The study of male infertility has become a priority for biologists and researchers in the last decades as a consequence of the declining birth rates. This problem has also become a major public health with economic and psychosocial consequences.

Analysis of human sperm cells, for instance, is widely used in investigations related to male infertility and assisted conception. Sperm samples are usually analysed by health professionals using microscope devices following a manual process to count and describe their morphology. Nevertheless, this practice is prone to errors and time consuming.
This thesis proposes a novel framework based on image processing and machine learning methods to automate the analysis of sperm cells. The proposed method presented an average accuracy performance of 96.4% classify automatically sperm cells in three classes: normal, abnormal and non sperm cell. Performance results have been obtained in challenging conditions: presence of uneven illumination, unwanted noise and blurring caused by the focus drift and occlusion of objects as a result of the overlapping of sperm cells, among others. The object of interest, sperm cells, captured in the images used in this research did not receive any staining or fixation treatment prior to their capture.

A novel and robust methodology based on deep neural learning is developed as part of the automatic feature selection prior to the classification. Also, video and image database of sperm samples was produced at the Andrology laboratory of the University of Sheffield as part of this work. The database was used to validate the proposed framework for the segmentation and classification of in-vitro cells.