Automatic segmentation of Sperm's parts in microscopic images of human semen smears using concatenated learning approaches

Accurate segmentation of the sperms in microscopic semen smear images is a prerequisite step in automatic sperm morphology analysis. It is a challenging task due to the non-uniform distribution of light in semen smear images, low contrast between sperm's tail and its surrounding region, the existence of various artifacts, high concentration of sperms and wide spectrum of the shapes of the sperm's parts. This paper proposes an automatic framework based on concatenated learning approaches to segment the external and internal parts of the sperms. The external parts of the sperms are segmented using two convolutional neural network (CNN) models which produce the probability maps of the head and the axial filament regions. To obtain acrosome and nucleus segments, the K-means clustering approach is applied to the head segments. A Support Vector Machine (SVM) classifier is used to classify each pixel of the axial filament segments to extract tail and mid-piece regions from obtained segments. The proposed method is validated on the images of the Gold-standard dataset. It achieves 0.90, 0.77, 0.77, 0.78, 0.75 and 0.64 of the average of dice similarity coefficient for the head, axial filament, acrosome, nucleus, tail, and mid-piece segments, respectively. Experimental results demonstrate that the proposed method outperforms state-of-the-art algorithms for the head and its internal parts segmentation. It also segments the axial filament region and its internal parts with desirable accuracy. Different from previous works, the proposed method is able to segment all parts of the sperms which enables automatic quantitative analysis of the sperm morphology.