NMN treatment reverses unique deep radiomic signature morphology of oocytes from aged mice

doi:10.21203/rs.3.rs-951364/v1

Download PDF

Research

NMN treatment reverses unique deep radiomic signature morphology of oocytes from aged mice

https://doi.org/10.21203/rs.3.rs-951364/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 28 Jun, 2022

Read the published version in Biomedicines →

Version 1

posted

You are reading this latest preprint version

One of the earliest physiological consequences of advancing age is the loss of female reproductive potential. This is primarily due to oocyte quality and developmental competence, which is highly sensitive to biological age. We applied deep learning, swarm intelligence and discriminative analysis to images of mouse oocytes taken by common bright field microscopy and were successfully able to identify a highly informative deep radiomic signature (DRS) of oocyte age. This signature distinguished morphological changes in oocytes associated with maternal age with 92% accuracy (AUC~1), reflecting this decline in oocyte quality. We then employed the DRS to evaluate the impact of the treatment of reproductively aged mice with the NAD⁺ precursor nicotinamide mononucleotide (NMN). The DRS signature classified 60% of oocytes from NMN-treated aged mice as having a 'young' morphology, suggesting that NMN was able to rejuvenate the morphological changes identified by the DRS. These findings indicate that NMN therapy may be able to restore the quality of a sizable subset of oocytes affected by reproductive ageing, and that these oocytes will be able to be distinguished and selected by DRS.

General Microbiology

Immunology

Oocyte

Aging

Morphology

Machine learning

NMN

In humans, the non-renewable reserve of oocytes is laid down during in utero development, where they must be maintained prior to ovulation and be ready for fertilisation decades later. Given the decades that oocytes must persist in the ovarian environment, oocytes are highly vulnerable to ageing, with impacts that occur well before other tissues[1]. Oocyte ageing represents a key constraint on natural and assisted reproduction, yet the ability to objectively measure the impacts of ageing on oocytes in a non-invasive manner is lacking. Currently, oocyte morphological analysis is undertaken using light microscopy by trained embryologists to provide a rough assessment of oocyte quality. However, visual assessment is highly subjective and poorly predictive [2]. Little work has been done on the impact of biological aging on oocyte morphology, with one study observing no difference for 20, 25 and 30 weeks of age in mice on zona pellucida thickness or periventricular space in mice[3]. An accurate methodology for assessing age-related differences in oocyte morphology is needed. This would have potential clinical utility (for optimizing oocyte selection) as well as enabling new avenues of research for reversing age-related declines in oocyte quality.

Recently, artificial intelligence (AI) has been widely applied to objective biomedical image assessment for disease diagnosis and monitoring in order to enable the precise customisation of treatment plans[4]. Deep learning strategies have been used to interpret electroencephalogram (EEG), electrocardiogram (ECG), magnetoencephalography (MEG), and magnetic resonance imaging (MRI) data, to improve reliability and precision[5]. This recent surge in interest has led to several attempts to apply AI methodologies to the assessment of embryo viability for human assisted reproduction, although success has been variable [6, 7]. AI methods have been proposed to automate sperm, embryo assessment through morphology analysis such as time laps imaging[8–12].

In this study, we hypothesized that ageing impacts oocyte morphological properties and generates a specific morphology signature in oocytes. As such, we applied a deep radiomic signature (DRS) method to detect age-related changes in oocyte morphology. DRS is an AI quantitative approach for automating the extraction of information from images to standardise the interpretation of medical imaging. Brightfield images of oocytes were segmented and morphological features were extracted using deep structured nets which capture features including shapes and textures. This approach identifies image differences resulting from characteristics perceptible to the human eye, such as oocyte size and circularity. In addition, this technique can capture characteristics that are imperceptible by traditional visual inspection of oocytes, including the specific spatial distribution of image pixel intensities and pixel interrelationships, where there are no existing mathematical definitions with which they can be captured[13].

To discover the morphological DRS for the impact of maternal age on oocyte morphology, we used a novel combination of artificial intelligence methods, including deep learning[14], swarm intelligence[15] and discriminative analysis[16]. Subsequent to DRS discovery, we applied our new DRS to brightfield microscopy images of oocytes obtained from aged mice that had been treated with nicotinamide mononucleotide (NMN). NMN is an orally deliverable metabolic precursor to the metabolic redox cofactor nicotinamide adenine dinucleotide (NAD+), which we have shown positively impacts female reproductive ageing [17]. In this way we tested our hypothesis that the DRS would be sensitive to reversals of the impact of ageing by comparing the morphology of old, NMN oocytes to that of young and old oocytes. To the best of our knowledge, this is the first study where age related oocyte morphological changes have been quantified using artificial intelligence.

1. Oocyte collection and NMN treatment

The oocyte images used in this study were brightfield microscopy images which were collected during a recent study [17] but whose morphology was not previously analyzed. To recover metaphase II (MII) oocytes, aged (12-month-old) and young (4- to 5-week-old) C57BL/6J female mice were maintained in individually ventilated cages at 22^oC at 80% humidity at a density of 5 per cage, with ad libitum access to food and water. All water in this animal house was acidified to pH 3 with HCl to decrease microbial growth. The UNSW animal house maintained a 12hr light/dark cycle with lights on at 0700 and off at 1900. Experiments were carried out with prior approval of the UNSW Animal Care and Ethics Committee (ACEC) under ACEC number 18/133A. UNSW ACEC operates under strict animal ethics guideline from the National Health and Medical Research Council (NHMRC) of Australia.

Aged females were treated with NMN in drinking water (2g/l) for 4 weeks. Oocytes from young mice are of very high quality and have high developmental competence, as such no additional benefit is seen for NMN treatment and hence this was not examined. After 4 weeks, both aged and young females were treated with an intraperitoneal (i.p.) injection of pregnant mare serum gonadotrophin (PMSG; Folligon, Intervet, Boxmeer, Holland) to stimulate follicle growth, followed by an i.p. injection of human chorionic gonadotrophin (hCG; Chorulon, MSD Animal Health, Australia 46 h later to induce ovulation. Young females were administered 5IU PMSG and 10IU hCG, whereas aged animals were treated with 10IU PMSG and 10IU hCG COCs were collected from oviductal ampullae using a 27-guage needle and collected in HEPES-buffered α-minimum essential medium (α-MEM; GIBCO Life Technologies, Grand Island, NY) supplemented with 3mg/ml bovine serum albumin (BSA; Sigma Aldrich. St Louis, MO) 14-16 h after hCG injection. Oocytes were stripped of their cumulus cells with hyaluronidase (concentration, supplier etc). Non-degenerate, nominally healthy oocytes were then placed into equilibrated Hank's balanced salt solution (HBSS) under paraffin oil for imaging. Finally, the number of collected oocyte from young, old and old- NMN treated mice are 26, 21, 29, respectively.

2. Microscopy Imaging

Standard brightfield microscopy imaging was performed on an Olympus iX83 system with a 40× oil objective (NA 1.15) and a Prime95B™ sCMOS (Photometrics) camera operated below -30°C to reduce noise. The image size was 1200×1200 pixels.

3. Data analysis

We applied a novel artificial intelligence strategy combining deep learning, swarm intelligence and discriminative analysis to images of mouse oocytes taken by bright field microscopy to create a highly informative deep radiomic signature (DRS) of oocyte morphology. The process of data analysis is illustrated in Figure 1. After bright field imaging, oocytes were segmented to isolate the image sections containing oocytes only. Then, oocyte images were augmented[18] to artificially expand the dataset by adding images that are intuitively equivalent to the original images (details in section 3.1. Data augmentation). Next, old and young oocyte were provided to the deep learning nets which were constructed to extract the deep information where several (here N=3) bespoke deep learning nets with different structures and resolutions were implemented to extract accurate, data-driven image feature information (details in section 3.2. Deep convolutional neural network). Further, the information associated with old and young oocytes was used to discover the DRS.

To discover DRS, we used discriminative analysis based on the feature subset iteratively selected from the pool of features by swarm intelligence, a technique that draws on collective behavior of a group of naïve agents[19] (details in section 3.3. Swarm intelligence). To this end, datapoints were partitioned to training data set (80% of data ) and testing data set (20% of data) through cross validation process. The DRS is able to take advantage of numerous image characteristics including size, circularity, spatial distribution of variations in pixel intensities and pixel interrelationships. Using training data set, feature values from the oocyte groups under consideration (here, young and old oocytes) are represented in a 2-D discriminative space spanned by two canonical variables providing the highest separation of these clusters measured by the Fisher distance (FD) (ratio of between-cluster and within-cluster variances)[20]. The canonical variables in our work are the optimal linear combinations of the utilized features. Further, the testing data points are projected onto this 2D discriminative space, and their Fisher distance (FD) is evaluated. This is followed by the next round of feature selection (new subset) by swarm intelligence and discriminative cluster analysis where the maximization of FD calculated on the testing data serves as the criterion in the swarm intelligence process. This iterative search for improved feature subsets is carried out until the algorithm achieves satisfactory convergence of FD with respect to its changes between consecutive swarm intelligence iterations. This DRS was then used to obtain the support vector machine classifier (details in section 3.4)) allowing to distinguish old vs young oocytes which were rigorously cross validated and finally we used that classifier to evaluate the NMN+old oocytes and produced the assessment of the NMN treatment outcomes.

3. 1. Data Augmentation

Image augmentation[18] is a technique to artificially expand the dataset by adding images that are intuitively equivalent to the original images, in this case images of oocytes that have been rotated by various angles, or reflected along various axes – as obviously the oocyte orientation on the microscope is irrelevant). This leads to enhancements in both the quantity and diversity of the data for training models, improving the performance and ability of the model to generalize. With image augmentation, CNN is able to learn features that are invariant with respect to their location in the image and image orientation. Image augmentation can aid the model in learning features that are invariant to intuitively acceptable image transforms such as left-to-right orientation to top-to-bottom ordering, etc.

In this work, we applied image augmentation to deal with the limited number of training images and enhance the performance of the CNN for discovering the DRS. For data augmentation, oocyte images and their mirror reflections were rotated at different angles (45^o, 90^o, 135^o, 180^o).

3.2. Deep Convolutional Neural Network

Convolutional neural network (CNN) is a deep learning strategy that automatically performs feature identification[21]. To identify features, the CNN carries out several procedures called convolutional layers that are sequentially applied to an input image to learn the features that in traditional algorithms were derived based on mathematical feature definitions[22]. This independence from mathematical definitions that represent prior knowledge in the feature extraction is the major advantage of CNN. A convolutional layer contains a number of sub-procedures called filters where image convolution with specific filter arrays is carried out to extract image information. As an example, a filter with 3x3 array whose parameters are all 1/9 computes the average of 9 pixels of the image after convolution. The number of filters can vary layer to layer. The parameters of these arrays could be taken from the literature, where one can source filter arrays from convolutional nets previously trained on large image data sets[23]. This is possible because some image features, such as edges, shapes, corners and intensity, are common in a wide range of images, enabling knowledge transfer [23]. Such filter sharing adds efficiency while maintaining good generalization[24]. Alternatively, the filters could be learned from the dataset through learning process[25]. Training is the step where the network learns from the data. Each filter array is assigned with random parameters and the classifier goes a forward pass based on the data to predict the class labels. Further, the predicted class labels are compared against the actual class labels and an error is calculated. This error is subsequently back propagated across the network and parameters are revised accordingly [25].

To apply the filter to the input image, the filter array is moved across the width and height of the input image and the dot products between the input image and filter array are computed at every spatial position. The output of the filter is another image of reduced size compared with the input image to the filter. Each convolutional layer is immediately followed by two other procedures called an activation and pooling layers. The specific activation layer used in this work referred to as a “rectified linear unit (ReLU)” removes negative pixels in the input image replacing them with zeros but retains all positive pixels. The role of the pooling layer is to reduce the spatial size of the input image by a chosen pooling operation. Here, we used max pooling where each 3x3 image tile is replaced by the maximum value in that tile. The activation and pooling layers lead to more effective training, by eliminating negative values, downsampling (making images smaller) and reducing the number of parameters that the network needs to learn.

The output of each convolutional layer is a modified image used as the input to the next convolutional layer. The convolution, ReLU and pooling processes are repeated until the final high-level information about the image (image features obtained through deep learning) is extracted at different resolution depending on the filters and the specifics of convolutional layers employed in the nets which alter the data with the intent of learning the features[24]. After learning the features by using several convolutional layers, the CNN typically shifts to classification through the next set of protocols called “fully connected layers” as in a standard multilayer neural network approach[25] however this leads to many unknown parameters that can only be defined through training on thousands of images[26], which is extremely challenging for clinical experiments and their limited number of images. Deep networks that only have a limited amount of training data suffer from reduction of accuracy and generality power of the model, especially when the depth (number of layers) of the network increases[27].

In this study, three CNNs (Net1, Net2 and Net3) were used to extract image features. Net 1 consisted of 152 convolutional with specific filters taken from ResNet[28]. The parameters in the first 151 convolutional layers were taken exactly as in ResNet and then Net 1 was trained using oocyte data to correct the deep features based on the actual oocyte features. Net 2 contained 22 convolutional layers. It was built by drawing on GoogLe net[29] where 21 layers used the filters from GoogLe and the 22-nd layer was fine-tuned by the training data set of the oocytes. Such fine-tuning corrects the net parameters to align them more closely with oocyte morphological features. Net 3 was developed specifically for this study. It had 5 convolutional layers, where the first 4 convolutional layers used the filters from Krizhevsky net[30]. The fifth convolutional layer had 96 filters with size 3x3 pixels; these filters were trained using the oocyte data set. In the end, by applying these nets, we were able to generate N_f~6000 of deep learning image features for our dataset. We did not use CNNs for feature classification, this was carried out by the method of swarm intelligence detailed below.

3.3. Swarm Intelligence

Swarm intelligence is a methodology inspired by the evolution of a group of simple information-processing interacting agents[15]. In this approach, the naive artificial agents (in this case feature subsets) are iteratively evolved according to a pre-set evolution rule attempting to find the highest FD as a criterion optimization problem[31]. First, a population of the agents (feature subsets) are generated, and the FD is assessed for each of these agents, and then the agents are repeatedly updated according to a defined evolutionary strategy until the convergence condition for FD is satisfied.

In this study, we chose the agents to be candidate feature subsets of all available features generated by our CNN (N_f~6000) and 50 agents was used. We have run the swarm intelligence process multiple (19) times to optimize the number of features K underpinning the deep radiomics signature, starting from K=2 to K = 20.

3.4. Support Vector Machine Classifier

In this work we used a support vector machine classifier (SVM), a strong supervised method that can deal with sparse data with limited risk of being overfitted. In this approach, a hyperplane is formed[32] with maximum margins in the high dimensional spectral feature space to separate data points into the classes under consideration (here, young and old oocytes). This classifier defines the data label based on a linear predictor function[33]. Then the classifier is trained using optimal DRS to predict the pre-defined data labels (here, old and young). To train our SVM classifier the method of 10 - fold cross validation was employed, and the classifier performance was evaluated using nested cross validation and bootstrapping[34].

Our approach obtained comprehensive morphological information from brightfield images of oocytes obtained following the super-ovulation of young (4-5 weeks) and reproductively aged (12 months) mice. Following the training of our DRS on these images of oocytes from young and aged animals, we applied this system to images of oocytes from a separate group of reproductively aged (12 months) mice that were treated with NMN (See Methods section 1). Figure 2 shows representative brightfield oocyte images of oocytes from these three groups of animals. All oocytes were non-degenerate and therefore nominally healthy in terms of reproductive competence, and were not morphologically distinct by visual inspection.

After applying image augmentation and a deep convolutional learning approach, the DRS was discovered using combination of the swarm intelligence method with discriminative analysis which was cross-validated using a testing dataset (Supplementary Figure 1, further information in Supplementary Material Section 1). The number of features in the swarm intelligence subsets (DRS dimension, Figure 3a) was independently varied and optimized and DRS was constructed using 15 features (30% from Krizhevsky net, 10% from Google net and 60% from Resnet). As shown in Figure 3b, this DRS allowed us to clearly separate clusters of young (red data points) and old oocytes (blue data points), highlighting a significant difference in morphology that could be automated for standardization of assessment. The final DRS based on 15 features was used to train our support vector machine classifier (furtehr details suplimentary material section 3). To this aim, 10 - fold cross validation was employed, and the classifier performance was evaluated using nested cross validation and bootstrapping (further information in Supplementary Material section. 2, Supplementary Figure 2-3). The corresponding receiver operating characteristic (ROC) graph determining the DRC classification performance (Figure 3c) showed it has an accuracy of 92.2±3.3 and area under the curve (AUC) of 0.99.

Brightfield images of oocytes from older, NMN treated mice from the recent work (Bertoldo et al. 2020) were taken concurrently with oocytes used to define the DRS (Figure 1) these oocytes were not used for the development of the canonical variables applied, and are fully independent. We showed that data points representing oocytes from NMN-treated aged mice (black crosses) moved away from the cluster of oocytes from untreated age-matched controls (blue circles) towards the young cluster (red squares; Figure 3d). Overall, ~55% of NMN-treated oocytes had morphological properties that exactly correlated to the young cluster, ~25% were very close to the young cluster and 20% retained the morphological properties of oocytes from old mice. Morphological features of oocytes from NMN-treated animals were subsequently fed to the support vector machine classifier trained with our DRS. This resulted in 60% of oocytes being sorted to the young group and 40% to the aged group. As NMN is known to restore 'youthful' characteristics and health to oocytes [17], this shown that the DRS is sensitive to changes induced in the biological age of oocytes induced by geroprotective interventions.

Morphology plays a significant role in developmental biology and is strongly affected by the cell’s microenvironment and response to biophysical and environmental factors[35]. Reliable oocyte morphology quantification indicative of oocyte biological (as opposed to chronological) age would be of significant utility to advise clinical decision making. However, there are no opportunities with sufficient consistency and precision for widespread uptake for oocyte morphology analysis. Conventional morphological analysis for oocyte competency scoring is mostly limited to visually visible features including that oocyte should have a clear, moderately granular cytoplasm, without inclusions, a small perivitelline space with a single unfragmented first polar body, and a round clear zona pellucida [2]. Generally, findings have been mixed with regards to whether such conventional scoring of oocyte morphology can be prognostic for oocyte quality assessment in terms of fertilization and [36]. A systematic review of fifty studies that investigated the impact of single or multiple oocyte features on in vitro fertilization (IVF) outcomes did not find any visible features with unanimous prognostic value for the further developmental competence of oocytes. More promising results were found for complex classification systems which considered multiple features. Extreme variability has also been observed between individual assessors and facilities applying scoring systems[37, 38]. Overall, while conventional assessment of oocyte morphology has demonstrated that morphology has potential to indicate the reproductive potential and quality of an oocyte[39], this promise has been difficult to realize through manual observations.

This study introduced a standard quantitative approach to assess oocyte morphology properties named DRS which enabled oocytes to be sensitively categorized according to their age category. We used brightfield images of oocytes from young and reproductively aged animals to extract morphological features and develop a DRS for the morphology of ageing in oocytes. In this study we demonstrated that a DRS is able to differentiate young and old oocytes with 92.2±3.3 accuracy. To extract features and create a comprehensive feature bank, three different nets with different structure were employed. We used extremely deep (Res net), deep (Google net) and moderately deep (Krizhevsky net) structure to capture features with different resolution. DRS discovered using swarm intelligence shows that features have been selected from all three nets. This demonstrates that the efficacy of using three networks while using only one net may result in suboptimal DRS and consequently lower classification performance.

Further, we analyzed NMN of oocyte morphology when it is used to treat aged animals. NMN treatment supports the generation of the redox cofactor nicotinamide adenine dinucleotide, which is essential to fundamental metabolic processes including glycolysis and the TCA cycle, and also acts as a substrate for proteins involved in DNA repair and epigenetic maintenance, such as members of the poly-ADP-ribose (PARP) and sirtuin family. We recently showed that NAD+ levels decline with age, impairing oocyte function, and that restoration of NAD+ levels through treatment with its metabolic precursor NMN could restore oocyte quality and functional fertility in aged animals [17]. As the availability of competent oocytes is a rate limiting factor on human reproduction their rejuvenation, and subsequent identification, of high-quality oocytes is an important goal for reproductive healthcare in the context of an ageing population. In this study we showed that in NMN treated aged animals, oocyte morphological properties were restored to those of young oocytes in 60% of cases. As well as adding to the growing evidence that NMN may address reproductive ageing in females[17], these findings have implications for research into oocyte quality and ageing.

Our results indicate that a combination of modern artificial intelligence methods of deep learning and swarm intelligence coupled with discriminative analysis produces a DRS capable of recognizing age-related changes in oocyte morphology. Our methodology outperform the conventional oocyte morphology analysis as it is automatic, objective, fast and can extract and consider specific features which are undetectable to human vision and also there is no particular mathematical definition for them. As well as the potential for optimizing oocyte selection in clinical practice, DRS could greatly accelerate the efficiency of research in oocyte quality and ageing. As the only reliable measures of oocyte quality are future outcomes (e.g. fertilization, blastocyst development, implantation and/or pregnancy), which are time and labour intensive, experimentation on interventions to optimise oocyte quality are severely curtailed. The direct measurement of apparent oocyte age given by our DRS would enable high throughput investigations of potential interventions to restore oocyte competence in the face of reproductive ageing. This study was performed by bright-field microscopy which is the plainest of all the optical microscopy illumination methods and involves only basic equipment. Bright-field imaging is routinely used in reproduction laboratories and therefore our methodology has high translatability.

This study was conducted as a proof of principle using images from mouse oocytes, and it will be necessary to evaluate this technology in images of human oocytes to test its relevance for clinical practice as an extension of this study. The evidence is stronger that it could be used to accelerate experimental research in mice, however additional validation in independent data sets is still required. Although DRS is fairly strong to deal with relatively small data set to extract strong morphological signature (backed up with statistical evidence in the manuscript), enhancing the number of samples in a data set can improve the strength of the discovered DRS. Furthermore, this study showed the possibility of extracting unique morphological age signature and was calibrated with known young and aged oocytes; therefore, to consider DRS as oocyte competency for potential successful pregnancy, follow up experiments will be required to demonstrate that the DRS is reliably associated with reproductive primary reproductive outcomes. The success rate of reproduction depends highly on oocyte quality and the current pregnancy rate per retrieved oocyte is estimated at 4.5%. Therefore DRS, as a potential non-invasive approach to predict human oocyte developmental potential, has promise to reliably improve the prediction of viability and blastocyst formation[40] for future clinical practice.

AI: Artificial Intelligence

AUC: Area under the curve

CNN: Convolutional neural network

DRS: Deep radiomic signature

FD: Fisher distance

hCG: human chorionic gonadotropin

i.p.: intraperitoneal

MII: Metaphase II

NAD: Nicotine adenine dinucleotide

NMN: Nicotinamide mononucleotide

PMSG: pregnant mare serum gonadotropin

ReLU: rectified linear unit

SVM: support vector machine

Acknowledgments

General: We acknowledge assistance from Dr Ayad G. Anwer

Funding: This work is partially supported by the Australian Research Council: CE140100003, and DP170101863 and by National Health and Medical Research Council: APP1139763, APP1117538, APP1127821, APP1122484, APP1103689 and APP1093643. The salaries of MJB and DG working in the lab of LEW and RBG were partly paid by contract research from Jumpstart Fertility to UNSW. We acknowledge assistance from Dr Ayad G. Anwer.

Author contributions: AH conceived of this project, performed artificial intelligence, swarm intelligence, discrimination analyses and biostatistics and prepared the manuscript, MJB assisted with data collection, contributed to project design and helped prepare the manuscript, SBM assisted with data collection and contributed to project design and helped to prepare the manuscript, JMC contributed to the project design, assisted with data collection and helped prepare the manuscript, DMG assisted with data collection and helped prepare the manuscript, WLL assisted with project design and helped prepare the manuscript, RBG assisted with project design and helped prepare the manuscript, LEW assisted with project design and helped prepare the manuscript, EMG assisted with project design and data analysis and helped to prepare the manuscript and supervised the project. All authors approved the final article.

Declaration of Interests: LEW is a co-founder, shareholder, director and advisor of Jumpstart Fertility Inc, which was founded to develop the use of NAD precursors in reproductive medicine. The salaries of MJB and DG working in the lab of LEW and RBG were partly paid by contract research from Jumpstart Fertility to UNSW. LEW is also an advisor and shareholder in EdenRoc Sciences (Metro Biotech NSW, Metro Biotech, Liberty Biosecurity), and in Life Biosciences LLC and its daughter companies (Jumpstart Fertility, Continuum Biosciences, Senolytic Therapeutics, Selphagy, Animal Biosciences). LEW is an advisor and shareholder in Intravital Pty Ltd. RBG and WLL are co-founders, shareholders and directors of Anlagen Pty. Ltd. which develops fertility diagnostics. Other authors report no conflict of interest.

Cimadomo D, et al. Impact of maternal age on oocyte and embryo competence. Front Endocrinol. 2018;9:327.
Setti AS, et al. Relationship between oocyte abnormal morphology and intracytoplasmic sperm injection outcomes: a meta-analysis. European Journal of Obstetrics Gynecology Reproductive Biology. 2011;159(2):364–70.
Mohammadzadeh M, et al. Influential effect of age on oocyte morphometry, fertilization rate and embryo development following IVF in mice. Middle East Fertility Society Journal. 2018;23(2):117–20.
Louis CM, et al. Review of computer vision application in in vitro fertilization: the application of deep learning-based computer vision technology in the world of IVF. J Assist Reprod Genet. 2021;38:1–13.
Craik A, He Y, Contreras-Vidal JL. Deep learning for electroencephalogram (EEG) classification tasks: a review. J Neural Eng. 2019;16(3):031001.
Patil SN, et al. Deep learning techniques for automatic classification and analysis of human in vitro fertilized (IVF) embryos. J Emerg Technol Innov Res. 2018;5(4):100–6.
VerMilyea M, et al. Development of an artificial intelligence-based assessment model for prediction of embryo viability using static images captured by optical light microscopy during IVF. Hum Reprod. 2020;35(4):770–84.
Khosravi P, et al. Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization. NPJ digital medicine. 2019;2(1):1–9.
Bormann CL, et al. Performance of a deep learning based neural network in the selection of human blastocysts for implantation. Elife. 2020;9:e55301.
Javadi S, Mirroshandel SA. A novel deep learning method for automatic assessment of human sperm images. Comput Biol Med. 2019;109:182–94.
Liao Q, et al. Development of deep learning algorithms for predicting blastocyst formation and quality by time-lapse monitoring. Communications biology. 2021;4(1):1–9.
Kandel ME, et al., Reproductive outcomes predicted by phase imaging with computational specificity of spermatozoon ultrastructure. Proceedings of the National Academy of Sciences, 2020. 117(31): p. 18302-18309.
Guo Y, et al. Deep learning for visual understanding: A review. Neurocomputing. 2016;187:27–48.
LeCun Y, Bengio Y, Hinton G. Deep learning nature. 2015;521(7553):436–44.
Kennedy J. Swarm intelligence. In: Handbook of nature-inspired and innovative computing. Springer; 2006. pp. 187–219.
Habibalahi A, et al. Optimized autofluorescence spectral signature for non-invasive diagnostics of ocular surface squamous neoplasia (OSSN). IEEE Access. 2019;7:141343–51.
Bertoldo MJ, et al. NAD+ repletion rescues female fertility during reproductive aging. Cell reports. 2020;30(6):1670–81.
Perez L, Wang J, The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621, 2017.
Garnier S, Gautrais J, Theraulaz G. The biological principles of swarm intelligence. Swarm intelligence. 2007;1(1):3–31.
Tsuda K, Kawanabe M, Müller K-R. Clustering with the Fisher score. Advances in Neural Information Processing Systems, 2003: p. 745–752.
Yoo H-J. Deep convolution neural networks in computer vision: a review. IEIE Transactions on Smart Processing Computing. 2015;4(1):35–43.
Campbell JM, et al. Non-destructive, label free identification of cell cycle phase in cancer cells by multispectral microscopy of autofluorescence. BMC Cancer. 2019;19(1):1–11.
Pan SJ, Yang Q. A survey on transfer learning. IEEE Transactions on knowledge data engineering. 2009;22(10):1345–59.
Cipolla R, Pentland A. Computer vision for human-machine interaction. Cambridge university press; 1998.
Rawat W, Wang Z. Deep convolutional neural networks for image classification: A comprehensive review. Neural computation. 2017;29(9):2352–449.
Faes L, et al. Automated deep learning design for medical image classification by health-care professionals with no coding experience: a feasibility study. The Lancet Digital Health. 2019;1(5):e232–42.
Too EC, et al. A comparative study of fine-tuning deep learning models for plant disease identification. Comput Electron Agric. 2019;161:272–9.
He K, et al. Deep residual learning for image recognition. in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
Szegedy C, et al. Going deeper with convolutions. in Proceedings of the IEEE conference on computer vision and pattern recognition.
Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems. 2012;25:1097–105.
Habibalahi A, et al. Non-invasive real-time imaging of reactive oxygen species (ROS) using auto-fluorescence multispectral imaging technique: A novel tool for redox biology. Redox Biol. 2020;34:101561.
Furey TS, et al. Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics. 2000;16(10):906–14.
Gosnell ME, et al. Computer-assisted cystoscopy diagnosis of bladder cancer. Elsevier.
Efron B, Tibshirani RJ. An introduction to the bootstrap. CRC press; 1994.
Bakal C, et al. Quantitative morphological signatures define local signaling networks regulating cell morphology. science. 2007;316(5832):1753–6.
Petersen CG, et al. Relationship between visualization of meiotic spindle in human oocytes and ICSI outcomes: a meta-analysis. Reprod Biomed Online. 2009;18(2):235–43.
Ashrafi M, et al. Effect of oocyte dysmorphisms on intracytoplasmic sperm injection cycle outcomes in normal ovarian responders. Journal of Obstetrics Gynaecology Research. 2015;41(12):1912–20.
Rienzi L, et al. Significance of metaphase II human oocyte morphology on ICSI outcome. Fertility sterility. 2008;90(5):1692–700.
Rienzi L, Vajta G, Ubaldi F. Predictive value of oocyte morphology in human IVF: a systematic review of the literature. Hum Reprod Update. 2011;17(1):34–45.
Raimundo JM, Cabrita P. Artificial intelligence at assisted reproductive technology. Procedia Computer Science. 2021;181:442–7.

Supplementarymaterialswithfigures.docx

Download PDF

Journal Publication

published 28 Jun, 2022

Read the published version in Biomedicines →

Version 1

posted

You are reading this latest preprint version

NMN treatment reverses unique deep radiomic signature morphology of oocytes from aged mice

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Materials And Methods

1. Oocyte collection and NMN treatment

2. Microscopy Imaging

3. Data analysis

3. 1. Data Augmentation

3.2. Deep Convolutional Neural Network

3.3. Swarm Intelligence

3.4. Support Vector Machine Classifier

Results

Discussion

Abbreviations

Declarations

References

Supplementary Files

Status:

Journal Publication

Version 1