Investigating Into Artificial Intelligence Trained Algorithms and Their Effectiveness in Addressing Female Infertility

We are part of a progressive world, a world that is trying to redefine many aspects of lives of an individual with the help of Aartificial intelligence, Machine learning and Deep learning. We today are witnessing an exponential surge in newer technologies such as Aartificial intelligence (AI) integrated with Internet of Things (IoT) sensors in the form of Wearable Technology which are increasingly used in multiple fields in health sciences as they are very effective in the real-time monitoring of the patients. These technologies are so much prevailing in the form of trained models in medical research that we frequent come across terms like Internet-of-Medial Things (IoMT) and Medical Artificial Intelligence while reading medical literature. Today, such domains of technology are being increasingly explored as they are unlocking the promising possibilities in health sciences. With the help of various technologies of AI techniques such as supervised, semi-supervised unsupervised and reinforcement learning, doctors across the globe are making suitable and timely decisions in treating and diagnosing the disease.

Female infertility has always been an issue and social stigma in India and it’s not only the matter of physical health of the female but also related to psychological health and the social status of the women. Women as well as families experiences stress due to countless visits to gynecologist and unavoidable treatment expenses. Now a day’s couples facing such problems are mostly referred to expensive treatments like IVF which on average can ranges from INR 1,00,000 to 3,50000 inclusive of medicines and testing procedures.

From Ovarian stimulation to AI in ART that is Assisted Reproductive Technology, there has been noteworthy innovation in the integration of AI and IoT in the Medical Labs. AI algorithms are expected to bring about a more calculated, computed, and standardized approach to the infertility problem today females are facing. As we are aware of the fact that AI deals with the impersonation of human intelligence through machine learning to gain knowledge regarding pattern recognition and effective, reliable decision making in cost effective way, thus exploring this domain can give insight to the problem of infertility in females. So an attempt is made to introduce a general outlook of the promising applications of AI to study reproduction issues in females and how trained AI models can be a game changer for those who are suffering from infertility.

This review aims to establish connection between female infertility problems with AI and discusses how an effort can be made to resolve such sensitive health issue, keeping in view the recent and past contribution of emerging AI technologies in health sciences.

In the past few decades, AI , ML and DL have evolved in the medical field, and are creating numerous opportunities for innovation and development in female health care. Various AI solutions for reproductive issues in females have recently been modeled to cure infertility.

Download figure

Following is the review summary of the study of AI, ML and DL trained medical models. The summary is divided into two broad categories only which relate to serious infertility issues in women and contribution of AI in them.

Various techniques AI approaches used for the prediction and analysis of PCOS

A methodology with image segmentation added between feature extraction and CNN was proposed by Palvi Soni et al. to eradicate unwanted data from ultrasound image and to detect the disease with high precision. With Convolutional Neural Network they tried to classify the ultrasound image to determine PCO class or nonPCO class. With the help to KNN (K-Nearest Neighbor Algorithm) classification was done for the severity of the disease whether disease is mild, moderate or at high level, according to which one can be treated for infertility. The proposed methodology is expected to detect PCOD disease in minimum execution time so as to save the time of doctors. ⁶

A system i-HOPE was proposed by Amsy Denny et al. for the timely discovery and prediction of PCOS.  With the help of unsupervised learning algorithm PCA (Principal Component Analysis) redundant data set was eliminated and contributing significant features were selected. Classification of PCOS was done using various linear and nonlinear ML algorithms such as Naïve Bayes classifier method, K-Nearest neighbor (KNN), Random Forest Classifier (RFC), Classification and Regression Trees (CART) etc. The medical data sets required was acquired through patient survey. Accuracy estimation of each of the algorithms was carried out and the achieved accuracy of the proposed model was RFC with an accuracy of 89.02%. ⁷

Dr.V.Krishnaveni outlined a four-step strategy for a prediction model to diagnose PCOS, which will assist the physicians to take early decisions for treatment on their patients. The paper suggests building proposed prediction model by applying Computing Intelligence and data mining techniques. Two important processes were suggested in building a predictive model (1) choosing the right criteria to solve the problem and (2) selecting the appropriate prediction method from the various Data mining techniques. The projected framework consist of collecting sample medical data set attributes related to PCOS from various related research papers, which might belong to five categories: Symptom, Medical history, Clinical signs, Scan findings and Blood test and Hormone analysis. After sample collection to eliminate redundant data feature selection is suggested to be applied on the medical data set, for which NP Hard and Heuristic algorithm approach is recommended. Then appropriate data mining method is selected, for which some classification techniques and algorithms discussed are ID3 Algorithm, Naïve Bayes Algorithm, Artificial Neural Networks (ANN) Algorithm, K nearest neighbour algorithm etc. Finally for assessing the quality of the model both internal and external validating testing is suggested.⁸

Kiruthika and Ramya et al. proposed an automatic recognition method for follicle. The method is a amalgamation of wavelet transform and k-means clustering. To quantify visual differences and to reduce computational complexity RGB image is color transformed to L*a*b* Color space. Discrete wavelet transform decomposes the ultrasound image and it removes the high frequency elements in the image without degeneracy. The approximation coefficient which is obtained after wavelet transformation is taken as input to k-means clustering. For further highlighting the features and to get desired information from the cluster k-means algorithm is used. Further Laplacian of Gaussian edge operator (LoG) is used to extract region of interest and edges of follicle. On the basis of the samples taken from the ultrasound images texture parameters along with intensity is suggested to be taken during clustering to reduce the classification error as projected clustering algorithm was build on intensity alone due to which interfollicular regions were also observed as follicles.⁹

A technique for automatic discovery of follicles by quantifying the multiple features of an ultrasound image of ovary was implemented with the help of hybridized PSO (Particle Swarm Optimization) algorithm and ANN. by O. R. Isah et al. The best features of the ultrasound image were selected by PSO algorithm to improve the efficiency of the system, to lower MSE (Mean Square Error). The best features selected by PSO algorithm decide the number of inputs necessary for training a specified framework of ANN. Using MLP (Multi-Layer Perceptron), a follicle detection algorithm which detects regions of interest and a hybridized PSO-MLP classification algorithm were developed. Classification rate and accuracy of the network was shown by confusion matrix, which on the basis of the ability of network to differentiate between follicles and non-follicles classes determines the high classification rate and accuracy. The algorithm was developed in Matlab R2015a and produced an accurateness of 98.3%.¹⁰

2. Various AI/ML techniques used in IVF clinic

Infertility in female of reproductive age has forced many desperate couples to opt for IVF, which is a medical modus operandi where an ovum of female is fertilized by male gametes in a laboratory dish that is elsewhere outside the body of female. Once embryo reaches a multiple cell stage in about 5 days, embryologist selects the embryo and loads it into a catheter so that doctor can place it in the uterus for further development. Pregnancy takes place when this embryo attaches to and embed itself into the uterus. Various AL and ML techniques can be incorporated in Assisted Reproductive Technology where sperm and an ovum are fertilized In vitro.

Gowramma G S et al. explained the method of medical IVF dataset visualization in graphical format to find out various relationships between the IVF dataset attributes and present them in a way which is easy for the physicians to perceive. To find the best relationships between the attributes they have used the concept of machine learning and python plotting libraries (Matplotlib and Seaborn) with the help of software named anaconda navigator. In Bar chart approach which is plotted for embryos transferred attribute v/s no of the live birth count helped in showing number of live birth occurrence. Similarly Scatter Plot plotted for Year of Treatment Vs Total number of live births conceived with the help of IVF showed that the number of years of treatment increases the number of live births.¹¹

I Tur-Kaspa et al. using AI approach proposed a multilayer perceptron deep learning based oocyte yield prediction framework competent of predicting the count of female gametocyte yield for a given COS (controlled ovarian stimulation) protocol, which can be used by infertility specialists to enhance IVF success. The study was designed around medical dataset which was collected between 2017-2021, This medical data set was additionally condensed to 769 cycles and was used as input for the AI framework. ¹²

Mikkel Fly Kragh et.al. presented a technical study on the evaluation and comparison of AI trained models for embryo selection. In order to evaluate and compare AI trained models two main performance metrics were discussed: Binary classification metrics and Model-wide metrics. Both Binary and Model-wide metrics were further categorized into series of performance metrics: Sensitivity, Accuracy (Proportion of Correct Prediction), Specificity, PPV, NPV, AU to measure the performance of AI trained models. Clinical usefulness was also discussed to check how functional the model would be in clinical exercise. Randomized controlled trials were suggested to eliminate selection bias and for fair comparisons. ¹³

Jared Geller et.al. have presented a study to create and evaluate a convolutional neural network that can calculate live birth rates using static ultrasound images from a IVF process. For the rationale of the study they examined two classification problems that is calculating whether an embryo will lead to a successful pregnancy or not and predicting the result of that pregnancy. DL techniques, including data augmentation was used decrease model variance and to strengthen inadequate medical dataset. Standard training and validation dataset split was used to shun model overfitting. Medical dataset was gathered from clinics in South Florida between 2016 and 2019. The main objective of the study was to develop an AI trained framework that take in input day 5 moved blastocysts and which outputs a variable showing whether the image taken will finally led to a live birth or not.¹⁴

Kevin Loewke et.al. collected data from various IVF clinics across the United States to introduce a generalized AI trained framework for grading embryos at the blastocyst phase. A series of Convolutional Neural Network (CNN) were used to classify the cropped embryo ultrasound images. Comparison of results of AI trained framework with manual morphology ranking is also done. The major performance metric employed for assessing the AI trained model is AuC (Area under Curve). Visualization technique and attribution algorithm illustrated that AI framework was able to focus on significant features and sensitivity. ¹⁵

STORK, a fully automated framework formulated on Google’s Inception model developed by Pegah Khosravi et.al was used to categorize embryo images based on blastocyst ranking with the help of deep neural networks. The main objective was to remove the manual subjectivity of the grading process. AI based DNN approach was utilized to decide on highest quality embryos using a huge set of human embryo time-lapse image without preprocessing on the input images. STORK effectively calculated blastocyst quality with an AUC of >0.98. To train the system transfer learning was used and best parameters set achieved from ImageNet database was used to initialize network parameter.¹⁶

Noritoshi Enatsu et. al. invented FiTTE, an AI based system which combined blastocyst images and clinical data to predict blastocyst viability. Two types of algorithms were used DL algorithms for image inputs and ML algorithms for non-image inputs .The system also provided visual explanation via gradient-based localization. Gred-CAM technique (gradient-weighted class activation method) was used to produce heat-map which provided physician to recognize areas of clinical value.¹⁷

Almost all of the AI techniques discussed above are addressing two main problems regarding infertility issue: addressing the problem and early diagnosis of PCOS in the women and enhancing the prediction accuracy of IVF treatment on the basis of images of blastocyst stage embryos. But most of the other significant facet which can be a cause of infertility like endometriosis, early detection of the blockage of fallopian tubes, defective implantation etc have been neglected by the researchers. Also most of the research study is done specific to the areas in which they are developed, so it can yield different results with less accuracy and reliability on those who have different characteristics from other geographical areas. Like most of the research papers included in literature survey belongs to Asia.

The grading assessment of embryo in IVF is purely subjective, and also the complex structure of embryo that is not possible for human eye to capture can largely affect the embryo selection, thus putting question mark on the successful outcome of IVF in the form of live births ¹⁸. With manual grading system embryologists are not able to detect chromosomal abnormalities which can result in miscarriage or lifetime genetic syndrome. Also IVF end result can be multiple pregnancies, which has many risks associated with it like premature delivery and AI- Improved Embryo Selection can make a huge change in IVF positive outcome. With the ability of learning and recognizing complex desired patterns ML algorithms can help embryologist in making precise decision regarding grading the embryos and hence selecting the most feasible embryo for transfer. AI algorithms will not alter the standard IVF process; it will only enhance the probability of single successful clinical pregnancy. Therefore more extensive simulation studies are needed to be done to generalize AI models

1. 1.cfm https www niehs nih govhealthtopicsconditionsrepro-healthindex, Last. (2022) Reviewed:.
   Google Scholar

1. 2.Pasquali R, Stener-Victorin E, Yildiz B O, Duleba A J, Hoeger K et al. (2011) PCOS Forum: research in polycystic ovary syndrome today and tomorrow. Clin Endocrinol (Oxf). 74(4), 424-33.
   Google Scholar

1. 3.Parasar P, Ozcan P, Terry K L. (2017) Endometriosis: Epidemiology, Diagnosis and Clinical Management. Curr Obstet Gynecol Rep. 6(1), 34-41.
   Google Scholar

1. 4.Jennings L K, Krywko D M. (2022) Pelvic Inflammatory Disease. [Updated. Available from:https://www.ncbi.nlm.nih.gov/books/NBK499959/
   Google Scholar

1. 5.Timeva T, Shterev A, Kyurkchiev S. (2014) Recurrent implantation failure: the role of the endometrium. , J Reprod Infertil 15(4), 173-83.
   Google Scholar

1. 6.Soni Palvi, Vashisht Sheveta. (2019) Image Segmentation for detecting Polycystic Ovarian Disease using Deep Neural Networks. , International Journal of Computer Sciences and Engineering 7, 10-26438.
   Google Scholar

1. 7.Denny Amsy, Raj Anita, Ashok Ashi, Ram Maneesh, George Remya. (2019) i-HOPE: Detection And Prediction System For Polycystic Ovary Syndrome (PCOS) Using Machine Learning Techniques. 673-678.
   Google Scholar

1. 8.Krishnaveni V.A Roadmap to a Clinical Prediction Model. , Ix, No. Ii, Pp, With Computational Intelligence for PCOS,” Vol, 177-185.
   Google Scholar

1. 9.Ramya M M, Kiruthika V. (2014) Automatic Segmentation of Ovarian Follicle Using K-Means Clustering.
   Google Scholar

1. 10.Omeiza Rabiu Isah, A D Usman, A M Tekanyi. (2017) A hybrid model of PSO algorithm and artificial neural network for automatic follicle classification.
   Google Scholar

1. 11.G S. (2021) Visual Analytics System for Evaluating IVF Data Predicting Live Birth”. , JRP 1, 50-63.
   Google Scholar

1. 12.Tur-Kaspa D E Fordham, Perl O, Tur-Kaspa T, Rosentraub S, Rosentraub D et al. (2022) Artificial intelligence (AI) to optimize ovarian stimulation for IVF: from stimulation protocols to total and mature oocyte yield. D.P Cohen, P-618 Human Reproduction, Volume 37, Issue Supplement_1 deac107.568, https://doi.org/10.1093/humrep/deac107.568 .
   Google Scholar

1. 13.M F Kragh, Karstoft H.Embryo selection with artificial intelligence: how to evaluate and compare methods?. , J AssistReprodGenet 38, 2021-10.
   Google Scholar

1. 14.Geller J, Collazo I, Pai R, Hendon N, Lokeshwar S D et al.An artificial intelligence-based algorithm for predicting pregnancy success using static images captured by optical light microscopy during intracytoplasmic sperm injection. , J Hum Reprod Sci 2021, 288-92.
   Google Scholar

1. 15.Loewke K, Cho J H, Brumar C D, Maeder-York P, Barash O et al. (2022) Characterization of an artificial intelligence model for ranking static images of blastocyst stage embryos. Fertil Steril. 117(3), 528-535.
   Google Scholar

1. 16.Khosravi P, Kazemi E, Zhan Q. (2019) Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization. , npjDigit. Med 2, 21-10.
   Google Scholar

1. 17.Enatsu N, Miyatsuka I, An L M. (2022) A novel system based on artificial intelligence for predicting blastocyst viability and visualizing the explanation. Reprod Med Biol. 10-1002.
   Google Scholar

1. 18.Kumar Yogesh, Talwar Pankaj, Shukla Aakriti. (2022) Understanding Understanding Embryo Grading Embryo Grading Academy of ART Little by little, one travels far.
   Google Scholar