Classification of Embryos Using Microscopic Pictures

Abstract

This study focuses on classifying the microscopic images of embryo by using DL neural networks including Mobilenet that captures the crucial patterns in the images and classify whether the Embryo is Good or Bad for In-Vitro Fertilization (IVF). The paper introduces an enhanced DL-based approach for recognizing patterns within images and adjusting each input image to meet specified normalization criteria, as normalization plays a crucial role in ensuring consistency, comparability across data sets. The dataset consists of 1079 embryos from which 239 microscopy images were utilized to validate the proposed method.

Introduction

I. INTRODUCTION

In IVF sector the improved safety and efficiency remains a top priority within reproductive medicine. Which is constantly developing new techniques. Choosing an appropriate human embryo during the fertilization process involves navigating a complex array of biological pathways, each responsive to environmental stimuli. For IVF treatments to be successful, precise, and non-intrusive monitoring is essential. While conventional wisdom has its strengths, it also has few flaws. Recent advances in computing power and technology These advancements in computing power and technology have paved the way for creative solutions to tackle these problems. addressing these obstacles, which could lead to higher success rates for IVF treatments. The goal of embryo culture in IVF is to replicate, with utmost accuracy, the environment in which an embryo would normally grow and develop. Medical professionals and scientists are starting to take an interest in Batch culturing is developing as a potential alternative to the traditional method of isolating embryos, which involves growing embryos in individual drops of culture media. Collective nurturing is a method that improves developmental outcomes by creating a Custom micro-environment: by placing numerous embryos housed in one well of a culture plate.

The microscopical examination is still the gold standard for gauging embryo growth in many IVF clinics. Scheduled microscopy examinations of Evaluation of embryo viability based on morphometric parameters and mitotic activity division, two important indicators of development, constitute this method. However, there are risks linked with this approach as it frequently requires briefly interrupting the incubation conditions for a brief period. to visualize the embryo regulated culture environment. An option that is only starting to take shape is time-lapse technology, which records extensive and continuous data regarding embryonic development growth in an incubator in an ongoing, non-invasive manner. The advantages of this technology have been driving its uptake in wealthy nations, while underdeveloped nations may be unable to afford or lack the necessary infrastructure to fully embrace it. Several factors, such as available resources, level of skill, the decision between traditional microscopy and time-lapse technology hinges on several factors, including available resources, clinician expertise, and the specific needs of each patient. Both approaches offer advantages and Limitations.

The initial stage of AI-powered embryo assessment involves constructing a robust training dataset. This dataset is meticulously built through two primary methods: manual segmentation by trained personnel or automated segmentation algorithms.  Following segmentation, the isolated embryo images undergo a rigorous curation process. Each image is meticulously labelled and categorized based on pre-defined morphological criteria established by embryologists. Finally, the DL model leverages established medical criteria to develop its embryo evaluation capabilities.

This transition from traditional microscopy to cutting-edge DL analysis represents a significant leap forward in IVF technology. A particular difficulty in this sector lies in evaluating embryos nurtured in groups. To address this, the training dataset incorporates time-lapse microscopy recordings capturing dynamic progression of co-cultured embryos. The ultimate objective is to achieve perfectly segmented and organized embryo data, which serves as the foundation for further advancements in DL-based embryo assessment.

II. RELATED WORK

A. Deep Learning-Based Prediction of IVF Implantation Success Using Time-Lapse Embryo Image Sequences of Humans from Days 3 and 5.

In this sector of assisted reproduction, researchers are continuously working to improve the success of IVF treatments. Historically, assessment protocols primarily focused on evaluating embryonic cell quality on the third-day post-fertilization. However, with advancements in the understanding of embryonic development, newer protocols have shifted towards assessing blastocyst quality on the fifth day, recognizing it as a more predictive indicator of implantation potential. The integration of AI systems into IVF procedures symbolizes a typical advancement towards improving outcomes. These AI systems offer the potential to uncover hidden relationships between various characteristics of embryos, thus aiding in more accurate prediction of implantation success. While existing AI systems have primarily focused on assessing single blastocyst images, increased awareness of the growing significance of analysing time-lapse image sequences to capture dynamic developmental changes. In this context, the paper under consideration proposes a novel approach that extends beyond static image analysis to predict embryo implantation outcomes from time-lapse image sequences. The proposed approach encompasses two distinct models: one evaluating embryos based on their attributes on the third-day post-fertilization, and the other assessing the same embryos using image sequences captured on the fifth day. Notably, the authors introduce a Data Length Scheduler (DLS) algorithm designed to accommodate variations in the lengths of blastocyst stage sequences, thus enhancing the robustness of their predictive models. The importance of this work lies in its comprehensive approach to leveraging AI for embryo assessment, encompassing both early and late-stage developmental attributes. With an impressive accuracy rate of 76.9%, the proposed system surpasses existing state-of-the-art methods by a notable margin of 6%. The integration of deep learning with time-lapse microscopy presents promising avenues for improving IVF processes, potentially leading to better outcomes for patients.

B. Automated and accurate cytoplasmic detection of human zygote: a convolutional neural network-based, resilient platform for picture segmentation.

The purpose of this study was to compare the accuracy and uniformity of zygote cytoplasm segmentation measures taken by CNN to those taken by knowledgeable embryologists. With the help of 550 zygote photos annotated by embryologists, we trained a CNN to automatically partition the cytoplasm of a human embryo. In addition, we have examined the reliability of cytoplasmic area estimations using CNN-segmented pictures vs those obtained by qualified human embryologists, also the accuracy, repeatability, and impact of various shared image characteristics. Additionally, when 8377 zygote images were utilized for regularity classification using CNN results, it was discovered that the high level of agreement between two embryologists (98 ± 1.02%), as well as between the embryologists and the CNN (97.75 ± 1.45%), when it came to the cytoplasmic area. Inside the CNN system, there existed a perfect agreement of 100%. Additionally, CNN measurement accuracy was unaffected by picture noise or brightness; however, very zygote-shaped irregularities decreased precision to 95% for both. A CNN demonstrated good performance in classifying both regularly and irregularly shaped zygotes (area under the curve: 0.874 ± 0.043) in the automatic segmentation of 8377 images used for classification. The zygotes' areas were measured at 9741.34 ± 7951.83 μm², and their shapes were classified as regular (65.37%), slightly irregular (15.1%), moderately irregular (9.1%), and highly irregular (9.2%). The accurate measurement of the zygote cytoplasmic area, the classification of form regularity, and the assistance with embryo evaluation were all made possible by the CNN system due to these findings.

C. Human embryo segmentation models based on u-nets are compared.

Clinical embryologists traditionally assess the quality of the human embryos generated by in vitro fertilization; this technique is laborious and susceptible to human mistakes. It is possible to apply AI techniques to grade time-lapse microscopy (TLM) photographs. Because the embryo's backdrop contains numerous artifacts that could mislead the grading algorithms, segmenting the embryo from the background of TLM pictures is a critical step for evaluating embryo quality. Deep learning-based automated approaches for segmenting images of human embryos (blastocysts) on day 5 were compared in this work. In comparison to our suggested model, four completely convolutional deep models were constructed using a blend of two loss functions and two optimizers based on gradient descent.

These models included U-Net and three of its versions. The experimental findings on the test set validated that our modified Dilated Inception U-Net model with Adam optimizer and Dice loss performed better than other U-Net variants with 98.68% accuracy, 97.52% precision, 99.20% Jaccard index, and 98.52% precision, respectively.

III. METHODOLOGY

The project involves six essential modules, each crucial for developing and evaluating a robust embryo classification and prediction model. Initially, Dataset Preparation organizes human embryo images into "train," "validation," and "test" directories, optimizing model training and evaluation. Following this, Data Augmentation techniques like rotation, shifting and flipping diversify the dataset, enhancing its robustness. Subsequently, Feature Extraction extracts critical image characteristics using the pre-trained MobileNet model across training, validation, and testing datasets. This examination offers invaluable understanding for further improvement of the model. Moving on, Model Architecture constructs personalized classification layers atop the extracted features, enhancing the model's predictive capabilities. Model Training utilizes the extracted features and labels, with multiple epochs and early stopping to prevent overfitting. Finally, Model Evaluation rigorously assesses the model's generalization using the test dataset, computing key metrics like loss and accuracy. Through these modules, the project aims to offer insightful perspectives on embryonic classification and prediction, with potential implications for reproductive health.

Conclusion

In conclusion, the proposed DL-based approach for segmenting day-3 and day-5 embryo pictures symbolizes a substantial advancement in AI-driven embryo evaluation. By employing deep learning methods for classification, it makes the classification much easy and helps for precise embryo data preparation. Despite encountering certain limitations inherent in the process, particularly in the classification of clear and blurry images, the findings from this research are far-reaching. The advancements in AI are particularly valuable for developing nations aiming to incorporate AI into their IVF programs. It offers a practical approach to improve embryo assessment, potentially leading to increased success rates in IVF treatments.

References

[1] Dimitriadis, I., C. L. Bormann, P. Thirumalaraju, M. Kanakasabapathy, R. Gupta, R. Pooniwala, I. Souter et al. \"Artificial intelligence-enabled system for embryo classification and selection based on image analysis.\" Fertility and sterility 111, no. 4 (2019): e21. [2] Shen, Xi, et al. \"The valuable reference of live birth rate in the single vitrified-warmed BB/BC/CB blastocyst transfer: the cleavage-stage embryo quality and embryo development speed.\" Frontiers in Physiology 11 (2020): 521758. [3] Mantziou, Veronika, et al. \"In vitro teratogenicity testing using a 3D, embryo-like gastruloid system.\" Reproductive Toxicology 105 (2021): 72-90. [4] Cimadomo, Danilo, et al. \"Inter-centre reliability in embryo grading across several IVF clinics is limited: implications for embryo selection.\" Reproductive BioMedicine Online 44.1 (2022): 39-48. [5] Giorgetti, Claude, et al. \"Implantation: Embryo score to predict implantation after in-vitro fertilization: based on 957 single embryo transfers.\" Human Reproduction 10.9 (1995): 2427-2431. [6] De Assin, R. Ruiz, et al. \"Comparison of methods to determine the assigned value in an external quality control programme for embryo evaluation.\" Reproductive biomedicine online 19.6 (2009): 824-829. [7] Chavez-Badiola, Alejandro, et al. \"Embryo Ranking Intelligent Classification Algorithm (ERICA): artificial intelligence clinical assistant predicting embryo ploidy and implantation.\" Reproductive biomedicine online 41.4 (2020): 585-593. [8] Montag, Markus, Bettina Toth, and Thomas Strowitzki. \"New approaches to embryo selection.\" Reproductive biomedicine online 27.5 (2013): 539-546. [9] Campbell, Alison, et al. \"Modelling a risk classification of aneuploidy in human embryos using non-invasive morphokinetics.\" Reproductive biomedicine online 26.5 (2013): 477-485. [10] Sruthi, P., Premkumar, L.” Attribute-based storage supporting secure de-duplication of encrypted data in cloud”,International Journal of Recent Technology and Engineering, 2019, 7(6), pp. 418–421. [11] Shirisha, N., Bhaskar, T., Kiran, A., Alankruthi, K.” Restaurant Recommender System Based on Sentiment Analysis”,2023 International Conference on Computer Communication and Informatics, ICCCI 2023, 2023. [12] Bhanu, J.S., Bigul, S.D., Prakash, A.” Agricultural internet of things using machine learning”,AIP Conference ProceedingsThis link is disabled., 2021, 2358, 080010. [13] Y Ambica, Dr.N.Subhash Chandra, A Cascaded Deep Network for Abnormal Region Extraction from MR Brain Images.International Journal of Grid and Distributed Computing.Vol. 13, No. 2, (2020), pp. 1554-1563.

Copyright

Copyright © 2024 Dr. P. Shruthi, Mr. Anil Kumar, A. Lokesh, A. Sai Kiran, E. Manohar, E. Pranay. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.