A novel deep learning method for automatic assessment of human sperm images

Highlights

• Automatic assessment of human sperms’ morphological deformities in head, acrosome, neck, tail, and vacuole.

• Designing an effective and low computational cost deep learning model.

• Ability to work on low resolution images and unstained sperms.

• High accuracy & Real-time processing time.

• Preparing a freely available dataset with 1,540 images.

Abstract

Sperm morphology analysis (SMA) is a very important factor in the diagnosis process of male infertility. This research proposes a novel deep learning algorithm for malformation detection of sperm morphology using human sperm cell images. Our proposed method detects and analyzes different parts of human sperms. First of all, we have prepared an image collection, called the MHSMA dataset, which can be used as a standard benchmark for future machine learning studies in this problem. This collection consists of 1,540 sperm images from 235 patients with male factor infertility. This unique dataset is freely available to the public. After applying data augmentation techniques, we have proposed a sampling method for fixing data imbalance. Then, we have designed a deep neural network architecture and trained it to detect morphological deformities in different parts of human sperm—head, acrosome, and vacuole. Our proposed method is one of the first algorithms that considers the acrosome. In addition, our method can work very well with non-stained and low-resolution images. Our experimental results on the proposed benchmark show the high accuracy of our deep learning algorithm for detection of morphological deformities from images. In these experiments, the proposed algorithm has achieved $F_{0.5}$ scores of $84.74\text{\%}$, $83.86\text{\%}$, and $94.65\text{\%}$ in acrosome, head, and vacuole abnormality detection, respectively. It should be noted that our algorithm achieves a better accuracy than existing state-of-the-art methods in acrosome and vacuole abnormality detection on the proposed benchmark. Also, our method works very fast. It can classify images in real-time, even on a mainstream laptop computer. This allows an embryologist to quickly decide whether or not the analyzed sperm should be selected.

Introduction

Infertility is the problem of almost 15% of couples. The lack of pregnancy after 12 months of intercourse without protection is defined as infertility. About 30 to 40 percent of infertile couples suffer from male factor abnormalities [20,43]. One of the problems in male factor infertility is spermatozoa morphology abnormalities, which may show teratozoospermia or oligoasthenoteratozoospermia.

The quality of spermatozoa is one of the most important parameters for oocyte fertilization and embryo quality. It is shown that abnormalities in sperm correlate with cleaving embryo morphology at later stages [7]. In other words, the shape of sperm is reflected by sperm development during spermatogenesis. As a result, problems in sperm maturation causes abnormalities in sperm morphology and the functionality of egg fertilization [2]. By assessment of sperm parameters and seminal plasma characteristics such as semen pH and sperm morphology, viscosity, concentration, and motility, male factor infertility can be determined [6].

The first successful live birth of a child using intra-cytoplasmic sperm injection (ICSI) method happened in 1992 [33]. These days, the ICSI method is widely used for the treatment of various couples who need assisted reproductive technologies. These couples can have normal, mildly, or severely abnormal semen parameters [5]. In several studies, the positive correlation between high ICSI outcomes and normal sperm morphology has been proven. In other words, serious abnormalities of the sperm head cause low fertilization, implantation, and pregnancy rates [30].

The normal sperm morphology is defined in previous studies by Menkveld et al. [28]. The length of a normal sperm head is between 3 and 5 micrometers. This range for the width of the sperm head is between 2 and 3 micrometers (two-thirds or three-fifths of head length). The length of midpiece is one and a half of the head length and it is axially presented $1\mu m$ in width. The normal tail should be also visible, uniform, uncoiled and thinner than the midpiece. Its length must be about $45\mu m$.

In the intracytoplasmic morphologically selected sperm injection (IMSI) procedure, sperm selection is performed at high magnification (usually 6000$\times$) [30]. However, existing microscopes of laboratories commonly have a low magnification (400 $\times$ and 600$\times$). These magnification levels are routinely utilized for sperm selection in the ICSI procedure. The visual assessment of sperms are also commonly performed manually and it is only based on the judgment of embryologists. This method is inexact, subjective, non-repeatable, and hard to teach. Another solution for the assessment of men fertility is computer-aided sperm analysis (CASA) using different staining procedures [6]. Due to the flaws of manual solutions, automatic techniques are essential for analyzing human sperm morphology. As a result, designing efficient and accurate algorithms for analyzing and classification of sperms and selecting the best one before ICSI is a challenging and trending task [38]. In other words, automatic methods for selection of the best sperm during the ICSI process and without staining will be more desirable for embryologists. This will lead to higher fertilization and pregnancy rates.

In this paper, we propose a novel deep learning algorithm for automatic extraction of sperm morphological features. This problem is a challenging task, due to the following reasons: 1) the number of sperm images is not enough for the training phase; 2) the normal and abnormal sperm classes are highly imbalanced, thus making the problem harder; 3) the pictures are taken using a low-magnification microscope and the details of these images are not clear; 4) the pictures are very noisy; 5) the sperms should not be stained; and 6) the analysis should be done in real-time to be useful for treatment purposes.

As the first step, we introduce a new dataset for our deep learning purpose. This dataset is based on the Human Sperm Morphology Analysis dataset (HSMA-DS) [15]. It is a unique dataset based on the number of images and its characteristics. For solving problems of training sample scarcity and class imbalance, we apply data augmentation and sampling techniques, respectively. Then, we propose a deep neural network architecture that can be trained to classify the normality of sperm acrosome, head, and vacuole. The proposed model has 24 convolution, two max pooling, one average pooling, and two fully-connected layers. Our proposed model, unlike a lot of existing methods, is able to work well with non-stained pictures. In other words, our method is applicable for treatment purposes. In addition, the time required for checking each sperm is very short (under 25 milliseconds) and our method can work in real-time. Our experimental results also show the effectiveness of our method, which can be regarded as the state-of-the-art for this dataset.

The rest of the paper is organized as follows: Some previous works are reviewed in Section 2. The prepared dataset is presented in Section 3. Our proposed algorithm, including the designed deep learning model, is depicted in Section 4. Sections 5 and 6 contain a comprehensive description of experimental setup, comparisons, and discussion. Finally, conclusions and future work are summarized in Section 7.