Artificial intelligence in human in vitro fertilization and embryology
2020; Elsevier BV; Volume: 114; Issue: 5 Linguagem: Inglês
10.1016/j.fertnstert.2020.09.157
ISSN1556-5653
AutoresNikica Zaninović, Zev Rosenwaks,
Tópico(s)Insurance, Mortality, Demography, Risk Management
ResumoEmbryo evaluation and selection embody the aggregate manifestation of the entire in vitro fertilization (IVF) process. It aims to choose the “best” embryos from the larger cohort of fertilized oocytes, the majority of which will be determined to be not viable either as a result of abnormal development or due to chromosomal imbalances. Indeed, it is generally acknowledged that even after embryo selection based on morphology, time-lapse microscopic photography, or embryo biopsy with preimplantation genetic testing, implantation rates in the human are difficult to predict. Our pursuit of enhancing embryo evaluation and selection, as well as increasing live birth rates, will require the adoption of novel technologies. Recently, several artificial intelligence (AI)-based methods have emerged as objective, standardized, and efficient tools for evaluating human embryos. Moreover, AI-based methods can be implemented for other clinical aspects of IVF, such as assessing patient reproductive potential and individualizing gonadotropin stimulation protocols. As AI has the capability to analyze “big” data, the ultimate goal will be to apply AI tools to the analysis of all embryological, clinical, and genetic data in an effort to provide patient-tailored treatments. In this chapter, we present an overview of existing AI technologies in reproductive medicine and envision their potential future applications in the field. Embryo evaluation and selection embody the aggregate manifestation of the entire in vitro fertilization (IVF) process. It aims to choose the “best” embryos from the larger cohort of fertilized oocytes, the majority of which will be determined to be not viable either as a result of abnormal development or due to chromosomal imbalances. Indeed, it is generally acknowledged that even after embryo selection based on morphology, time-lapse microscopic photography, or embryo biopsy with preimplantation genetic testing, implantation rates in the human are difficult to predict. Our pursuit of enhancing embryo evaluation and selection, as well as increasing live birth rates, will require the adoption of novel technologies. Recently, several artificial intelligence (AI)-based methods have emerged as objective, standardized, and efficient tools for evaluating human embryos. Moreover, AI-based methods can be implemented for other clinical aspects of IVF, such as assessing patient reproductive potential and individualizing gonadotropin stimulation protocols. As AI has the capability to analyze “big” data, the ultimate goal will be to apply AI tools to the analysis of all embryological, clinical, and genetic data in an effort to provide patient-tailored treatments. In this chapter, we present an overview of existing AI technologies in reproductive medicine and envision their potential future applications in the field. Discuss: You can discuss this article with its authors and other readers at https://www.fertstertdialog.com/posts/31478 Discuss: You can discuss this article with its authors and other readers at https://www.fertstertdialog.com/posts/31478 Embryo evaluation and selection embody the aggregate manifestation of the entire in vitro fertilization (IVF) process. It aims to choose the “best” embryos from the larger cohort of fertilized oocytes, the majority of which will be determined to be not viable either as a result of abnormal development or due to chromosomal imbalances. Indeed, it is generally acknowledged that even after embryo selection based on morphology, time-lapse microscopic (TLM) photography, or embryo biopsy with preimplantation genetic testing (PGT), implantation rates in the human are difficult to predict. While contemporary embryo evaluation methods have generally improved, the IVF process has remained gamete inefficient and embryo wasting. Ideally, methods that enable selection of both healthy sperm and oocytes will optimize the IVF process and gamete efficiency. Coupled with the necessity to fertilize fewer oocytes, novel noninvasive embryo assessment methods will increase IVF efficiency and reduce embryo wastage. Assessing human oocytes in a clinical setting is primarily done by observing the morphological appearance of the oocyte cumulus complex (1Zaninovic N. Elemento O. Rosenwaks Z. Artificial intelligence: its applications in reproductive medicine and the assisted reproductive technologies.Fertil Steril. 2019; 112: 28-30Abstract Full Text Full Text PDF PubMed Scopus (16) Google Scholar). While the initial assessment of corona cells may be indicative of oocyte maturity, it is only after cumulus removal and observation of the first polar body that oocyte nuclear maturity is confirmed. Ideally, noninvasive methods that identify markers of cytoplasmic maturity need to be developed. Applying machine learning (ML) to oocyte images before intracytoplasmic sperm injection (ICSI), as well as assessing oocyte behavior during ICSI, could become crucial for selecting competent oocytes. For example, the competency of mouse oocytes matured in vitro has been evaluated and has been predicted with high accuracy using an artificial neuronal network on TLM images (2Cavalera F. Zanoni M. Merico V. Bui T.T.H. Belli M. Fassina L. et al.A neural network-based identification of developmentally competent or incompetent mouse fully-grown oocytes.J Vis Exp. 2018; 133: 56668Google Scholar). Identifying noninvasive markers to evaluate oocyte competency may also be useful in research on alternative methods of oocyte creation, for example, oocyte in vitro maturation, nuclear cloning, and stem cell as well as direct somatic cell reprogramming. Some early attempts have already been made with artificial intelligence (AI) methods to evaluate human oocytes and predict normal fertilization, assess embryo development to the blastocyst (BL) stage, and even analyze implantation potential using static oocyte images. These early results need to be confirmed with larger studies (ESHRE 2020, O-285, U.S. Patent no. 2020/012623). Artificial intelligence has also been applied in semen analyses (3Agarwal A. Henkel R. Huang C.C. Lee M.S. Automation of human semen analysis using a novel artificial intelligence optical microscopic technology.Andrologia. 2019; 51e13440Crossref Scopus (15) Google Scholar) to evaluate sperm morphology (4Javadi S. Mirroshandel S.A. A novel deep learning method for automatic assessment of human sperm images.Comput Biol Med. 2019; 109: 182-194Crossref Scopus (27) Google Scholar) and DNA integrity (5McCallum C. Riordon J. Wang Y.H. Kong T. You J.B. Sanner S. et al.Deep learning-based selection of human sperm with high DNA integrity.Commun Biol. 2019; 2: 250Crossref Scopus (21) Google Scholar) as well as for sperm selection. A popular sperm analysis method called the CASA integrated a low-level AI ML system for automatic sperm evaluation (6Goodson S.G. White S. Stevans A.M. Bhat S. Kao C.Y. Jaworski S. et al.CASAnova: a multiclass support vector machine model for the classification of human sperm motility patterns.Biol Reprod. 2017; 97: 698-708Crossref Scopus (20) Google Scholar). A major target for clinical AI application is the identification of sperm cells in microsurgical testicular samples of patients with severe male factor infertility, as identifying these “precious” cells typically requires several hours by embryologists. Developing such a system will require a massive number of sperm images for machine training in order to correctly differentiate sperm from other tissue cells. The system will also need to be rapid and applicable in real time. Attempts to apply AI in urology based on a patient’s clinical parameters have already been reported: that is, AI methods have been developed to predict the likelihood of sperm extraction in azoospermic patients (7Zeadna A. Khateeb N. Rokach L. Lior Y. Har-Vardi I. Harlev A. et al.Prediction of sperm extraction in non-obstructive azoospermia patients: a machine-learning perspective.Hum Reprod. 2020; 35: 1505-1514Crossref Scopus (14) Google Scholar) and to reliably predict male factor infertility (8Chu K.Y. Nassau D.E. Arora H. Lokeshwar S.D. Madhusoodanan V. Ramasamy R. Artificial intelligence in reproductive urology.Curr Urol Rep. 2019; 20: 52Crossref Scopus (8) Google Scholar). The use of data-mining methods to determine the impact of lifestyle and environmental factors on seminal quality and fertility rates in men has also been explored (9Sahoo A.J. Kumar Y. Seminal quality prediction using data mining methods.Technol Health Care. 2014; 22: 531-545Crossref PubMed Scopus (30) Google Scholar). The application of AI technologies has gone even further with the development of smartphone-based applications for semen analysis as well as sperm viability and DNA integrity (10Kanakasabapathy M. Thirumalaraju P. Yogesh V. Natarajan V. Bormann C.L. Bhowmick P. et al.Automated smartphone-based system for semen assessment through the hyaluronic binding assay.Fertil Steril. 2017; 108: E74Google Scholar, 11Dimitriadis I. Bormann C.L. Kanakasabapathy M.K. Thirumalaraju P. Kandula H. Yogesh V. et al.Automated smartphone-based system for measuring sperm viability, DNA fragmentation, and hyaluronic binding assay score.PLoS One. 2019; 14e0212562Crossref Scopus (11) Google Scholar). Imaging is one of the most significant areas of AI application. Artificial intelligence technology has been successfully applied to identify objects within an image and predict shapes and textures. In medicine, it has been applied widely for image recognition and prediction in pathology, diagnostic radiology, and ultrasound (US), to mention a few. The current focus of AI applications in embryology can be categorized into the following groups: automatic annotation of embryo development (cell stages and cell cycles), embryo grading (mostly in the BL stage), and embryo selection for implantation. The emergence of TLM in human embryology has enabled the precise evaluation of the timing of cellular divisions and the detection of normal and abnormal hallmarks of embryo development. Even with moderate acceptance in the United States, due to high equipment costs and the policy of many clinics to use PGT as the standard of care, this technology can serve to standardize embryo culture systems and optimize embryo assessment. Indeed, applying AI to TLM will result in the reemergence of this technology. Commercially available TLM systems—EmbryoScope (Vitrolife), Geri (GeneaBiomedix), and ESCO—claim that their evaluation software includes some level of ML. These manufacturers have not disclosed their platform nor the performance accuracy of their systems. Furthermore, most automatic annotation and embryo selection programs are not available in the United States, as they are not Food and Drug Administration approved. During TLM incubation, embryologists can annotate the precise time of each cleavage event. However, this is a manual process that depends on each embryologist’s precision, experience, and ability to distinguish between normal and abnormal cleavage events and features (fragments vs. cells). As with any manual task, intra- and interoperator variability, particularly between different laboratories, is high (12Sundvall L. Ingerslev H.J. Breth Knudsen U. Kirkegaard K. Inter- and intra-observer variability of time-lapse annotations.Hum Reprod. 2013; 28: 3215-3221Crossref PubMed Scopus (90) Google Scholar). For this reason, along with the aim of standardizing and automatizing the process, AI-based automatic annotations with high accuracy and reliability are essential. There have been attempts to automatize TLM annotation. Apart from commercial manufacturers of TLM instruments, recent publications have demonstrated the success of automatic, non-human-mediated embryo developmental annotation systems (13Dirvanauskas D. Maskeliunas R. Raudonis V. Damasevicius R. Embryo development stage prediction algorithm for automated time lapse incubators.Comput Meth ProgBiomed. 2019; 177: 161-174Crossref PubMed Scopus (17) Google Scholar, 14Malmsten J. Zaninovic N. Zhan Q. Rosenwaks Z. Shan J. Automated cell division classification in early mouse and human embryos using convolutional neural networks.Neural Comput Appl. 2020; https://doi.org/10.1007/s00521-020-05127-8Crossref Scopus (3) Google Scholar, 15Raudonis V. Paulauskaite-Taraseviciene A. Sutiene K. Jonaitis D. Towards the automation of early-stage human embryo development detection.Biomed Eng Online. 2019; 18: 120Crossref Scopus (13) Google Scholar, 16Feyeux M. Reignier A. Mocaer M. Lammers J. Meistermann D. Barriere P. et al.Development of automated annotation software for human embryo morphokinetics.Hum Reprod. 2020; 35: 557-564Crossref Scopus (8) Google Scholar). Automatic annotation systems must comply with the following requirements: they should be fast, accurate, reproducible, and specific (i.e., able to distinguish a cell from a fragment), and they should recognize abnormal cellular developments (i.e., direct unequal cleavages) (17Zhan Q. Ye Z. Clarke R. Rosenwaks Z. Zaninovic N. Direct unequal cleavages: embryo developmental competence, genetic constitution and clinical outcome.PLoS One. 2016; 11e0166398Crossref PubMed Scopus (60) Google Scholar). Additionally, they should include the ability to distinguish morphological features of the embryo (uneven size, vacuoles, granularity, etc.) as well as nuclear abnormalities (multinuclear blastomeres). The ideal system would require that the weight (importance) of each characteristic be properly assigned and calculated. There are different computerized image systems available, such as cell shape extractors, segmentation, cell tracking, and feature extraction, that can be combined with AI systems (mostly with convolutional neural network [CNN] methods). Applying automatic detection technology to analyze cell stages and corresponding division intervals (t times) using TLM systems presents a challenge. First, the system needs to recognize and detect an embryo in the culture well and create an automatic region of interest. This can be done by a cascade classifier (15Raudonis V. Paulauskaite-Taraseviciene A. Sutiene K. Jonaitis D. Towards the automation of early-stage human embryo development detection.Biomed Eng Online. 2019; 18: 120Crossref Scopus (13) Google Scholar) or segmentation (18Giusti A. Coranil G. Gambardella L. Magli C. Gianaroli L. 3D localization of pronuclei of human zygotes using textures from multiple focal planes.Med Image Comput Comput Assist Interv. 2010; 13: 488-495Google Scholar). A recent publication from our group indicates that preprocessing of the region of interest is not necessary if the Inception V3 CNN model is used for automatic cell annotation. We believe that embryo segmentation is performed internally by the AI algorithm. We achieved an accuracy of 100% on mouse images within two frames and 93.9% accuracy on human TLM images within five frames, up to the eight-cell stage. The high success of the algorithm can be attributed to using up to five time frames for exact cell prediction and using different focal planes available with the TLM system (14Malmsten J. Zaninovic N. Zhan Q. Rosenwaks Z. Shan J. Automated cell division classification in early mouse and human embryos using convolutional neural networks.Neural Comput Appl. 2020; https://doi.org/10.1007/s00521-020-05127-8Crossref Scopus (3) Google Scholar). Feyeux et al. (16Feyeux M. Reignier A. Mocaer M. Lammers J. Meistermann D. Barriere P. et al.Development of automated annotation software for human embryo morphokinetics.Hum Reprod. 2020; 35: 557-564Crossref Scopus (8) Google Scholar) recently achieved automatic annotation up to the BL stage by using a segmentation tool to quantify zona pellucida thickness for identifying BL expansion. Additionally, the authors used the standard deviation of image gray levels to accurately identify embryo cell stages. The automated tool was validated with embryologists’ annotations and showed differences depending on cell-stage assessment (16Feyeux M. Reignier A. Mocaer M. Lammers J. Meistermann D. Barriere P. et al.Development of automated annotation software for human embryo morphokinetics.Hum Reprod. 2020; 35: 557-564Crossref Scopus (8) Google Scholar). Most advances in AI embryology have been made in embryo grading, specifically on BLs. The BL stage is particularly suited for grading, as it has been shown to have a significant association with implantation. Unfortunately, there are many grading systems. Even with the universally used “Gardner” system, variations and deviations have been quite common. The biggest problem is that this grading system uses a combination of numbers and letters rather than numerical values. Recently, we attempted to solve this problem by assigning a simple numerical BL score that incorporates qualitative numerical values for all three BL components: expansion, the inner cell mass (ICM), and the trophectoderm (TE) (19Zhan Q. Sierra E.T. Malmsten J. Ye Z. Rosenwaks Z. Zaninovic N. The blastocyst score, blastocyst quality ranking tool, is a predictor of blastocyst ploidy and implantation potential.Fertil Steril Rep. 2020; 1: P133-P141Google Scholar). There are more significant problems with BL evaluation and selection, namely that high intra- and interoperator variabilities exist (20Richardson A. Brearley S. Ahitan S. Chamberlain S. Davey T. Zujovic L. et al.A clinically useful simplified blastocyst grading system.Reprod BioMed Online. 2015; 31: 523-530Abstract Full Text Full Text PDF PubMed Scopus (22) Google Scholar, 21Khosravi P. Kazemi E. Zhan Q. Malmsten J.E. Toschi M. Zisimopoulos P. et al.Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization.NPJ Digit Med. 2019; 2: 21Crossref PubMed Scopus (133) Google Scholar), even when using TLM (11Dimitriadis I. Bormann C.L. Kanakasabapathy M.K. Thirumalaraju P. Kandula H. Yogesh V. et al.Automated smartphone-based system for measuring sperm viability, DNA fragmentation, and hyaluronic binding assay score.PLoS One. 2019; 14e0212562Crossref Scopus (11) Google Scholar). A recent study of 10 embryologists showed a very high degree of variability in grading human embryos for biopsy and freezing. In contrast, trained AI-based algorithms outperformed human evaluation and selection (22Bormann C.L. Thirumalaraju P. Kanakasabapathy M.K. Kandula H. Souter I. Dimitriadis I. et al.Consistency and objectivity of automated embryo assessments using deep neural networks.Fertil Steril. 2020; 113: 781-787.e1Abstract Full Text Full Text PDF Scopus (27) Google Scholar). One of the biggest challenges in this kind of work is the quality of the training data and how the machine learns from it. For example, the machine learns from a set of embryo images (training set) that have been evaluated and graded by embryologists (human involvement). How well the machine learns depends on the quality of the data. It would be ideal if ML could occur without human involvement. Early attempts to address this issue focused on differentiating the ICM from the TE using image segmentation. This required the use of two different focal images: one on the ICM and the other on the TE. The system used “simple” ML methods (support vector machine) on two-dimensional BL images. Although these attempts required human intervention, they represent the first step toward BL grade standardization and automatization (23Santos Filho E. Noble J.A. Poli M. Griffiths T. Emerson G. Wells D. A method for semi-automatic grading of human blastocyst microscope images.Hum Reprod. 2012; 27: 2641-2648Crossref Scopus (68) Google Scholar). The automatic evaluation of the BL is also an attractive model for animal research and animal IVF. By applying an artificial neural network using specific measurements, textures, and other BL features, Matos et al. (24Matos F.D. Rocha J.C. Nogueira M.F. A method using artificial neural networks to morphologically assess mouse blastocyst quality.J Anim Sci Technol. 2014; 56: 15Crossref Google Scholar) demonstrated that an automatic morphological classification of mouse BLs can be achieved with 95% accuracy. This group successfully applied its AI algorithm to bovine embryos as well, which was a prerequisite for applying AI methods to the evaluation of human embryo images (25Rocha J.C. Passalia F.J. Matos F.D. Takahashi M.B. Ciniciato D.S. Maserati M.P. et al.A method based on artificial intelligence to fully automatize the evaluation of bovine blastocyst images.Sci Rep. 2017; 7: 7659Crossref PubMed Scopus (34) Google Scholar). Several papers have described the identification and segmentation of the ICM from the TE on BL images using various AI tools (26Saeedi P. Yee D. Au J. Havelock J. Automatic identification of human blastocyst components via texture.IEEE Trans Biomed Eng. 2017; 64: 2968-2978Crossref Scopus (22) Google Scholar). Efforts were made to segment, evaluate, and predict TE quality in order to accurately and automatically predict BL quality (27Rad R.M. Saeedi P. Au J. Havelock J. Trophectoderm segmentation in human embryo images via inceptioned U-Net.Med Image Anal. 2020; 62101612Abstract Full Text Full Text PDF PubMed Scopus (16) Google Scholar, 28Singh A. Au J. Saeedi P. Havelock J. Automatic segmentation of trophectoderm in microscopic images of human blastocysts.IEEE Trans Biomed Eng. 2015; 62: 382-393Crossref PubMed Scopus (28) Google Scholar). It would also be of interest to analyze the use of AI methods to evaluate and select TE cells for embryo biopsy and compare them to the chromosomal results. Various attempts have been made to replicate BL morphological grading by automatic systems. These aimed to predict the ICM and TE grade automatically by using static or TLM images. One approach included training TLM BL data graded by embryologists on the ICM and TE. The preprocessing of the images was done using cropping and a CNN combined with a recurrent neural network to predict the ICM and TE morphology from the same images within three focal planes. The overall AUC was between 0.63 and 0.65 for the ICM and TE, with a high accuracy of distinguishing A (high) versus C (low) grade ICM and TE (97.8% and 98.1%, respectively) (29Kragh M.F. Rimestad J. Berntsen J. Karstoft H. Automatic grading of human blastocysts from time-lapse imaging.Comput Biol Med. 2019; 115103494Crossref PubMed Scopus (31) Google Scholar). Another paper has described the ability of a CNN to predict BL expansion (96% accuracy), ICM (91% accuracy), and TE (84% accuracy) quality grades using single static images taken under an inverted microscope (30Chen T.-J. Zheng W.-L. Liu C.-H. Huang I. Lai H.-H. Liu M. Using deep learning with large dataset of microscope images to develop an automated embryo grading system.Fertil Reprod. 2019; 1: 51-56Crossref Google Scholar). Although an AI system can be applied to analyze static images, the most significant benefit will be its use with TLM videos (stacked images) of developing embryos. There is some debate surrounding the use of static images versus videos for embryo assessment. On the one hand, it seems logical that videos would provide more information about embryo development, leading to better selection. On the other hand, the examination of a single embryo image at critical time points (i.e., at 66 or 110 hours) appears to be sufficient to precisely assess its developmental competence. The drawback of this system is that the single image cannot be used to evaluate the dynamic process of embryo development or events, such as blastocoel collapse, which can potentially interfere with the assessment. These two models need to be evaluated on the same data sets to draw valuable conclusions (31Manna C. Nanni L. Lumini A. Pappalardo S. Artificial intelligence techniques for embryo and oocyte classification.Reprod BioMed Online. 2013; 26: 42-49Abstract Full Text Full Text PDF PubMed Scopus (57) Google Scholar). In addition, does the quality of the image pixel resolution influence the accuracy of an AI system? Is it possible to train an AI system on high-resolution images and apply it to low-resolution images, or vice versa? Our recent study (Khosravi et al. (21Khosravi P. Kazemi E. Zhan Q. Malmsten J.E. Toschi M. Zisimopoulos P. et al.Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization.NPJ Digit Med. 2019; 2: 21Crossref PubMed Scopus (133) Google Scholar)) opened the possibility of applying AI technologies clinically to the classification of human embryos. Our analysis was performed on a large number of TLM images at the BL stage at exactly 110 hours post-ICSI. That specific time was determined based on our experience evaluating BLs using TLM. At that time (the morning of day 5), competent known implanted embryos exhibit morphological characteristics associated with a high BL grade. Interestingly, other authors independently used a similar time for BL evaluation (32Thirumalaraju P. Kanakasabapathy M.K. Bormann C.L. Gupta R. Pooniwala R. Kandula H. et al.Evaluation of deep convolutional neural networks in classifying human embryo images based on their morphological quality.arXiv. 2020; arxiv.org/abs/2005.10912Google Scholar). Generally, as we have learned from TLM, the speed and timing of embryonic developmental events are strongly associated with implantation. It is now well established that a BL developed by day 5 has a higher implantation potential compared with a BL developed by day 6, even when normalized for BL quality and chromosomal normality by PGT (33Irani M. O'Neill C. Palermo G.D. Xu K. Zhang C. Qin X. et al.Blastocyst development rate influences implantation and live birth rates of similarly graded euploid blastocysts.Fertil Steril. 2018; 110: 95-102.e1Abstract Full Text Full Text PDF PubMed Scopus (42) Google Scholar). Using images, we developed an AI framework (called STORK) to classify human BLs with >98% accuracy when distinguishing high- and low-grade BLs. By using class-activating mapping, a method that generates heat maps of images, it appears that STORK’s decision-making relies on scrutinizing distinct areas within the embryo, depending on its developmental stage. Additionally, we generated a decision tree integrating BL quality, evaluated by STORK, and maternal age, creating a predictive tool for determining the likelihood of achieving pregnancy. A similar attempt was reported recently in which CNN and Xception architecture were used to assess TLM and static embryo images to predict BL versus non-BL embryos at 113 hours with high accuracy (32Thirumalaraju P. Kanakasabapathy M.K. Bormann C.L. Gupta R. Pooniwala R. Kandula H. et al.Evaluation of deep convolutional neural networks in classifying human embryo images based on their morphological quality.arXiv. 2020; arxiv.org/abs/2005.10912Google Scholar) as well as the potential for BL implantation (34Bormann C.L. Kanakasabapathy M.K. Thirumalaraju P. Gupta R. Pooniwala R. Kandula H. et al.Performance of a deep learning based neural network in the selection of human blastocysts for implantation.Elife. 2020; 9Crossref Google Scholar). Likewise, recently, novel embryo parameters were incorporated into AI architecture in efforts to improve the selection of BL that are likely to implant (35Bori L. Paya E. Alegre L. Viloria T.A. Remohi J.A. Naranjo V. et al.Novel and conventional embryo parameters as input data for artificial neural networks: an artificial intelligence model applied for prediction of the implantation potential.Fertil Steril. 2020; https://doi.org/10.1016/j.fertnstert.2020.08.023Abstract Full Text Full Text PDF Scopus (19) Google Scholar). As has been described in semen evaluation, a smartphone-based system with portable optical devices to evaluate human embryos has recently been developed. For this study, the training process used high-quality embryo images, and the learning algorithm was applied to low-quality images taken with a phone. The overall accuracy of distinguishing BLs versus non-BLs was approximately 90%, indicating AI adaptivity and application flexibility (36Kanakasabapathy M.K. Thirumalaraju P. Bormann C.L. Kandula H. Dimitriadis I. Souter I. et al.Development and evaluation of inexpensive automated deep learning-based imaging systems for embryology.Lab Chip. 2019; 19: 4139-4145Crossref Google Scholar). Several AI startups are currently developing algorithms focused on various stages of embryonic development. Some are focused on early developmental stages, including pronuclear formation and behavior, while others target times after day 3 and blastulation time. All methods aim to predict embryo development and implantation. The question is, why is there a focus on the different developmental periods? Additionally, many are claiming high predictive accuracy by using different AI platforms: CNN, deep learning, computer vision, and others. Certain AI platforms are better depending on the data used; for example, CNN is ideal for image analysis. These different approaches indicate that AI technology for IVF has not yet been fully optimized. Moreover, many published analyses have been based on relatively small input training data (<300) and lacked sufficient data diversity or were unbalanced. One of the major challenges in ML is using an unbalanced data set for training. If data are not balanced (i.e., an approximately equal number of positive and negative), the AI will tend to recognize data that are in the majority with higher probability. Selecting the best (most competent) single embryo for transfer is the quintessential goal of all IVF embryologists. Typically, contemporary embryo selection methods rely on morphological assessment, a method that has been reported to be associated with high interoperator variability and inconsistency. Artificial intelligence represents one of the most promising, objective tools for embryo selection and pregnancy prediction. Indeed, some group
Referência(s)