Comparative Analysis of Conventional and Modified Capsule Networks for Sperm Morphology Assessment

The evaluation of sperm morphology is an essential factor in the diagnosis of male-related infertility. Manual sperm morphology assessments are labor intensive and can vary significantly between observers. Although automated sperm morphology analysis approaches have shown significant success in the last two decades, there is still a need for deep learning based solutions due to the challenges faced in new data sets such as the dramatic increase seen in number of sperm samples and the diversity of the labeled abnormality types. As a solution to these challenges, this research performs a comparative analysis of the conventional capsule network (CapsNet) and its modified version (FixCaps), using a recently introduced imbalanced dataset referred to as Hi-LabSpermMorpho which contains a total of 18,456 images labeled into 18 different classes. The classes include variants of abnormalities of the head, neck, and tail of the sperm, as well as the normal class. The evaluation focused on assessing the accuracy, computational efficiency, and the effects of different optimization strategies for both models. The results have shown that FixCaps achieved higher performance than CapsNet, with its improvements such as using large-kernel convolutions and the CBAM attention mechanism, which allow the model to capture more complex features and spatial relationships of the images in the dataset. FixCaps resulted in 52.67% classification accuracy while CapsNet performed only 37.89%.