Table 4 Sperm assessment studies using artificial intelligence
From: The prospect of artificial intelligence to personalize assisted reproductive technology
Study | ART process | Outcomes of interest | Dataset | AI methods | Results |
|---|---|---|---|---|---|
Hicks et al. (2019)45 | Motility assessment | -Sperm video sequences used to predict motility in terms of progressive, non-progressive, and immotile spermatozoa. -Combined with participant data in multimodal analysis for automated prediction of motility parameters. | VISEM—live spermatozoa videos from 85 different participants. | Deep learning—CNN | -Deep learning algorithms capable of predicting sperm motility efficiently and with reproducibility. -Combination with participant clinical information did not improve prediction. -Incorporation of temporal analysis outperformed traditional machine learning approach. -Best MAE achieved with CNN was 8.74. |
Thambawita et al. (2019)46 | Motility assessment | -Extraction of temporal features from sequential frames from videos are able to predict motility and train traditional CNN models. | VISEM | Deep learning—CNN | -Increase in number of stacked video frames from 9 to 18 improves motility prediction, implying this model has capabilities to learn temporal features from different video frames. -Best MAE achieved with CNN was 8.74. |
Ottl et al. (2022)47 | Motility assessment | -Automatic sperm motility assessment using framework of unsupervised spermatozoa tracking, feature extraction, and ML. | VISEM | Linear Support Vector Regression | -Able to predict the percentage of progressive, non-progressive, and immotile spermatozoa in a given sample. -MAE reduced to 7.31; an improvement compared to previous papers. |
Saiffe Farías et al. (2022)48* | Motility assessment | -Individual operator-assessed single sperm morphology linked to motility patterns assessed by vision-based AI software in ICSI ready sperm. | 2154 individual sperm video recordings. | Vision-based AI Software SiD1 (IVF2.0 Ltd.) | -Spermatozoa classified as morphologically normal showed better motility variables (higher linear movement, straight line velocity). -Sperm tail morphology defects had the most significant impact on motility variables. -AI-driven sperm motility assessment may be sufficient to assess morphological features for sperm selection. |
Mendizabal-Ruiz et al. (2022)49 | Motility assessment | -Vision-based AI software assessing progressive motility parameters (straight-line velocity, linearity of curvilinear path, head movement patterns) to predict successful fertilization and blastocyst formation. | 383 individual spermatozoa videos from 78 ICSI cycles. | Vision-based AI Software SiD1 (IVF2.0 Ltd.) | -Statistically significant differences in progressive motility patterns measured by SiD1 between successful and unsuccessful fertilization, and blastocyst formation. -Possible avenue for carrying out real-time analysis of individual spermatozoa during selection for ICSI. |
Shaker et al. (2017)52 | Morphology assessment | -Sperm images labeled with a class were divided into patches to identify important features in the sperm. -Dictionary learning is more effective for sperm head classification than previously published shape-based feature recognition. -Developed HuSHeM dataset with consensus classification and freely available for research purposes. | HuSHeM—includes 216 sperm head images (54 normal, 53 tapered, 57 pyriform, and 52 amorphous). SCIAN-MorphoSpermGS—includes 1862 images of sperm shapes (100 normal, 228 tapered, 76 pyriform, 73 small, and 656 amorphous), partial consensus among three experts. | Traditional ML with adaptive patch dictionary learning | -62% accuracy with SCIAN-MorphoSpermGS dataset. -92.3% accuracy, 93.5% precision, and 92.3% recall with new HuSHeM dataset. |
Javadi and Mirroshandel (2019)56 | Morphology assessment | -Deep CNN trained to detect morphological deformities in head, acrosome, and vacuole. -Developed MHSMA dataset labeled with normal sperm (acrosome, head, vacuole, tail, and neck). | MHSMA—includes 1,540 sperm images from 235 subjects with male factor infertility. | Deep learning—CNN | -High accuracy for detection of morphological deformities in sperm acrosome, head, and vacuole. -Accuracy scores of 76.7%, 77%, and 91.3% in acrosome, head and vacuole abnormality respectively, which requires improvement. -Able to classify images in real-time, aiding in selection of sperm for ICSI. |
Abbasi et al. (2021)57 | Morphology assessment | -Deep CNN algorithms trained to detect morphological deformities in head, acrosome, and vacuole. | MHSMA | Deep learning—CNN | -AI models capable of predicting sperm head features more accurately than previous study (84%, 80.7%, and 94% for sperm head, acrosome, and vacuole respectively). |
Jiang et al. (2022)58* | Morphology and viability assessment | -Deep learning AI technique to predict viability of immotile sperm through morphology assessment with a single bright-field image. | 1471 images of immotile sperm from 15 semen samples for training 10 new semen samples for validation. | Deep learning—CNN | -AI model able to accurately predict sperm viability in non-invasive manner without sample processing or staining. -Subtle morphological changes to sperm nucleus detected by AI otherwise challenging to identify with the naked eye. -Yet to be externally validated. |
Joshi et al. (2023)124 | Morphology assessment | -Deep neural network for morphological classification of sperm sample videos captured at 40x objective magnification. | -32 cryopreserved donor semen samples with known teratozoospermia and 720 vitrified sibling-oocytes from donors. -Oocytes split evenly between two conditions: (1) standard ICSI performed according to laboratory protocols and (2) AI-assisted sperm selection prior to injection. | Deep learning—CNN | -AI-ICSI group resulted in relatively increased fertilization by 6.42% and blastocyst rate by 21.35%. -Formation of high quality blastocysts increased by 41.7% compared to standard embryologist selection. |
McCallum et al. (2019)62 | DNA fragmentation | -Correlation between spermatozoa image and DNA integrity from single bright-field image. | 1064 images of stained sperm with known DNA integrity. | Deep learning—CNN | -Deep CNN trained to predict DNA integrity from single spermatozoa image in under 10 ms. |
Kuroda et al. (2022)63* | DNA fragmentation | -Modified AI-aided sperm chromatin dispersion (SCD) counting device compared to conventional Halosperm G2 Test. | 17 semen samples | AI-driven SCD Kit | -AI-aided automatic counting device capable of determining DNA fragmentation quicker in a much larger sample (mean 500 spermatozoa analyzed manually in 20 min. vs. 2600 spermatozoa analyzed automatically in 5 min.), and with good correlation to conventional testing. -Automated AI device ‘X12’ had good correlation to conventional Halosperm G2 test (r = 0.69, p = 0.02), as well as the group’s modified SCD R10 manual test (r = 0.88, p < 0.01). |
Peng et al. (2023)64 | DNA fragmentation index | -ML-based clustering used to determine the effect of DNA fragmentation index and conventional semen analysis parameters on IVF outcomes. | 1258 fresh IVF cycles with DNA fragmentation index data. | Unsupervised k-means clustering | -Favorable IVF outcomes seen with low sperm DNA fragmentation values, in combination with high or moderate motility sperm concentration and motility levels. -Worst outcomes seen with high sperm DNA fragmentation values and low sperm motility and concentration levels (live birth odds ratio 0.62; 95% CI 0.39–0.97). |
