Human Blastocyst Classification after In Vitro Fertilization Using Deep Learning
Ali Akbar Septiandri, Ade Jamal, Pritta Ameilia Iffanolida, Oki Riayati, Budi Wiweko
HHuman Blastocyst Classification after In VitroFertilization Using Deep Learning
Ali Akbar Septiandri , Ade Jamal Faculty of Science and TechnologyUniversitas Al Azhar IndonesiaJakarta, Indonesia [email protected], [email protected] Pritta Ameilia Iffanolida, Oki Riayati, Budi Wiweko
Human Reproductive, Infertility, and Family Planning Research Center IMERIFaculty of MedicineUniversitas IndonesiaJakarta, Indonesia
Abstract —Embryo quality assessment after in vitro fertiliza-tion (IVF) is primarily done visually by embryologists. Variabilityamong assessors, however, remains one of the main causes ofthe low success rate of IVF. This study aims to develop anautomated embryo assessment based on a deep learning model.This study includes a total of 1084 images from 1226 embryos.The images were captured by an inverted microscope at day3 after fertilization. The images were labelled based on Veeckcriteria that differentiate embryos to grade 1 to 5 based on thesize of the blastomere and the grade of fragmentation. Our deeplearning grading results were compared to the grading resultsfrom trained embryologists to evaluate the model performance.Our best model from fine-tuning a pre-trained ResNet50 on thedataset results in 91.79% accuracy. The model presented couldbe developed into an automated embryo assessment method inpoint-of-care settings.
Index Terms —in vitro fertilization, embryo grading, deeplearning
I. I
NTRODUCTION
Embryo quality plays a pivotal role in a successful IVFcycle. Embryologists assess embryo quality from the morpho-logical appearance using direct visualization [1]. There arethree protocols in different time points to evaluate the qualityof an embryo: (1) quality assessment of zygote (16-18 hoursafter oocyte insemination), (2) morphological quality assess-ment of cleavage stage embryos (day 3 after insemination),and (3) Morphological quality assessment of blastocyst stageembryos (4-5 days after fertilization) [2].This kind of visual assessment is susceptible to the subjec-tivity of the embryologists. There are two kinds of variabilityin embryo assessment as seen in [3]: interobserver and in-traobserver. Grading systems like Veeck criteria [4] aim tostandardize grading and minimize both variabilities. However,as also found in [3], “the embryologists often gave an embryoa score different than Dr. Veeck, but that score was typicallywithin one grade.” The study also shows that the intraobservervariation is limited. Khosravi et al. [5] also shows that only89 out of 394 embryos were classified as the same quality byall five embryologists in their study.In recent years, we have applications of deep learning forcomputer vision in medical imaging to address this variabilityissue. From MRI for brain imaging [6], various anatomical areas [7], to point-of-care diagnostics from microscopy images[8], deep learning has aided medical practitioners to diagnosebetter. In the field of reproductive medicine, we have alsoseen some application of artificial intelligence as shown in[9]. Recent studies also explored the possibilities to automateembryo assessments for IVF [5], [10], [11] using a robustclassifier trained on thousands of images.Our main contribution is that previous studies [5], [10], [11]are using day 5 embryo images, while we are using day 3embryo images. As seen in [12], “Early embryos can grow in asimple salt solution, whereas they require more complex mediaafter they reach the 8-cell stage.” Day 3 and day 5 embryos aresimilar in implantation, clinical pregnancy, twinning, and livebirth rates [12], [13], but since day 5 embryos are extremelysensitive to suboptimal culture environment, many clinics arestill doing day 3 embryo transfers [14]. Moreover, unlikeprevious studies which only used ResNet50, we also comparedseveral deep learning architectures to do the task. The datasetused in this study is described in the following section.II. D
ATASET
Our dataset comprises of 1084 images from 1226 embryosof 246 IVF cycles at Yasmin IVF Clinic, Jakarta, Indonesia.The images were captured by an inverted microscope at day 3after fertilization. The images consist of 2-3 embryos each. Wemanually cropped the images and a team of 4 embryologistsgraded them 1-5 by using Veeck criteria [4]. However, weonly found grade 1 to grade 3 embryos from the samples.This yields 1226 identified embryos consisting of 459 grade1, 620 grade 2, and 147 grade 3 embryos.To train the model and for further generalization error, wedivided the dataset into train and test sets with 75:25 ratio. Thetraining set is further divided into training and validation setswith 70:30 ratio. Some examples of the images in the trainingset can be seen in Figure 1. These images were automaticallycropped and resized by the library that we used for the deeplearning application [15]. This preprocessing step was done toprevent overfitting of the models. a r X i v : . [ ee ss . I V ] A ug ig. 1. Embryo images after the preprocessing steps III. M
ETHODOLOGY
We used the fast.ai library [15] for the deep learningapplication. The library helped us to do transfer learningfrom several pre-trained convolutional neural networks [16],such as the residual networks [17] with different depths(ResNet18, ResNet34, ResNet50, ResNet101), densely con-nected convolutional networks [18], Xception [19], and theMobileNetV2 [20] to the given task. We trained these modelsusing backpropagation with 1cycle policy [21]. We used thecyclical learning rates [21] implemented in the fast.ai libraryto find the best learning rates. To find an unbiased estimate ofthe accuracy from a model, we trained the model and predictedthe test set five times after ensuring that we got the best modelduring the training step.
A. Deep Residual Networks
Prior to the study on residual learning, deeper neuralnetworks are harder to train [17]. The accuracy of deeperneural networks gets saturated and becomes worse eventually.Their solution is to recast the original mapping H ( x ) into F ( x ) + x that can be seen as a feedforward neural networkwith shortcut connections. This reformulation makes it easierto optimize the model and can even reach over 1000 layerswith no optimization difficulty though achieved worse resultcompared to the ones with fewer layers. The ensemble of thisarchitecture has 3.57% top-5 error on the ImageNet test setand also won the 1st places in several tracks in ILSVRC andCOCO 2015 competitions. B. Densely Connected Convolutional Networks
Borrowing ideas from deep residual networks, the DenseConvolutional Network (DenseNet) tries to harness the powerof shortcut connections. “For each layer, the feature-maps ofall preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers.” [18] This enables the model to have “substantially fewer parametersand less computation” while still achieving state-of-the-artperformances. Nevertheless, DenseNet can still be scaled intohundreds of layers easily.
C. Xception
We also used the Xception architecture [19] to benchmarkagainst the study in [10]. This architecture is an improvementfrom the Inception V3 [22] “where Inception modules havebeen replaced with depthwise separable convolutions”. Xcep-tion also uses residual connections as in the deep residualnetworks [17]. It outperforms Inception V3 while having asimilar number of parameters as Inception V3.
D. MobileNetV2
Since the result would possibly be implemented in a mobilephone for better outreach, we consider MobileNetV2 as aviable architecture. This architecture allows us to “reducethe memory footprint needed during inference by never fullymaterializing large intermediate tensors.” [20]IV. R
ESULTS AND D ISCUSSION
We found that the best learning rates are around × − from 8 epochs. After 8 epochs, the learning curve startsplateauing which suggest the model overfits the training data.The results from all models can be seen in Table I. We cansee that the ResNet50 got the highest accuracy and the lowestcross-entropy loss of all models. While increasing the depth ofthe ResNet model from 18, 34, to 50 increased the accuracy,it stopped increasing afterwards. We argue that this is becausethe dataset is relatively simpler compared to ImageNet wherewe have different objects in different colours and sizes. Thismight also be the case why DenseNet models failed to achievebetter performance while being more complex. TABLE IM
ODEL COMPARISON model accuracy lossResNet18 . ± .
75% 0 . ± . ResNet34 . ± .
27% 0 . ± . ResNet50 . % ± . % . ± . ResNet101 . ± .
00% 0 . ± . DenseNet121 . ± .
27% 0 . ± . DenseNet169 . ± .
54% 0 . ± . Xception . ± .
96% 0 . ± . MobileNetV2 . ± .
84% 0 . ± . On the other hand, we are more interested in MobileNetV2which achieved a similar accuracy to the best model. As weelaborated in the previous section, this architecture wouldenable us to design an embedded system for point-of-carediagnostics. Thus, we provide the learning curve from Mo-bileNetV2 in Figure 2. An example of a confusion matrixfrom MobileNetV2’s prediction can be seen in Figure 3.Note that we are only using three embryo grades in thisstudy due to the unavailability of the samples. We would needto reassess these models when we have more grades. However,ince the current model can predict the minority class (grade3) well, we argue that it might generalise with the completegrades as well.
Fig. 2. Learning curve from MobileNetV2Fig. 3. A confusion matrix from MobileNetV2
Examples from the misclassified embryos as seen in Fig-ure 4 suggest that the image capturing process can impact themodel performance. For example, different shades of color(bottom right image) of the embryo images might cause themisclassification. Obstruction in the images, such as the redcircles (top right image) or timestamps from the applicationthat we used to digitally process the microscopy images alsoaffect the performance of the models.V. R
ELATED W ORK
Robust assessment and selection of human blastocysts afterIVF using deep learning has been studied in [5], [10], [11].Khosravi et al. [5] fine-tuned their Inception V1 model [23]on 10,148 images of day 5 embryos from the Center for Re-productive Medicine at Weill Cornell Medicine. The accuracyof the resulting model is 96.94%. However, the training wasdone for good and poor quality images only from the threeclasses defined first.Kragh et al. [10] address the issue in [5] by predictingblastocyst grades of any embryo based on raw time-lapseimage sequences. Aside from using Xception [19] as the
Fig. 4. Examples of misclassified images convolutional neural network (CNN) architecture to extractimage features, they also use a recurrent neural network(RNN) “that connects subsequent frames from the sequencein order to leverage temporal information.” They predict theinner cell mass (ICM) and trophectoderm (TE) grades for theentire sequence from the RNN using two independent fully-connected layers. They train the models on 6957 embryos. Ona test set of 55 embryos annotated by multiple embryologists,their models reached 71.9% and 76.4% of ICM and TEaccuracy respectively compared to human embryologists whoonly achieved 65.1% and 73.8%. While the result is promising,using a RNN makes the training slower and prone to vanishingor exploding gradients [24].In [11], the authors fine-tune a ResNet50 model on 171,239images from 16,201 day 5 embryos to predict blastocyst de-velopment ranking from 36, ICM quality, and TE quality. Theimages were annotated by embryologists based on Gardner’sgrading system. They achieved “an average predictive accu-racy of 75.36% for the all three grading categories: 96.24% forblastocyst development, 91.07% for ICM quality, and 84.42%for TE quality.” VI. C
ONCLUSIONS
We have shown in this study that we can grade day 3 embryoimages automatically with the best accuracy of 91.79% byfine-tuning a ResNet50 model. We found that more complexmodels failed to achieve better accuracy compared to theResNet50. MobileNetV2 as our model of interest to buildan embedded system achieved a relatively similar accuracyof 91.14% compared to the best model. The models still facesome problems from different shades of colour or obstructionsfrom the software that embryologists use to capture andprocess the images.e saw good results when combining CNN and RNN inprevious studies. However, in resource constrained settings,e.g. when we want to make inferences on small devices, theunparallelizable nature of RNNs also makes it challenging toimplement. Moreover, time-lapse microscopes are not preva-lent in developing countries. Thus, our solution would be morefeasible to put into production.Since we are still manually cropping the embryos from theoriginal images, we can extend this work to automate thistask, e.g. using an image segmentation model like YOLOv3[25] or U-Net [26]. In the future, we hope that this modelcan be developed as an embedded system for point-of-carediagnostics such as found in [8].R
EFERENCES[1] J. Cummins, T. Breen, K. Harrison, J. Shaw, L. Wilson, and J. Hen-nessey, “A formula for scoring human embryo growth rates in in vitrofertilization: its value in predicting pregnancy and in comparison withvisual estimates of embryo quality,”
Journal of In Vitro Fertilization andEmbryo Transfer , vol. 3, no. 5, pp. 284–295, 1986.[2] N. Nasiri and P. Eftekhari-Yazdi, “An overview of the available methodsfor morphological scoring of pre-implantation embryos in in vitrofertilization,”
Cell Journal (Yakhteh) , vol. 16, no. 4, p. 392, 2015.[3] A. E. B. Bendus, J. F. Mayer, S. K. Shipley, and W. H. Catherino,“Interobserver and intraobserver variation in day 3 embryo grading,”
Fertility and sterility , vol. 86, no. 6, pp. 1608–1615, 2006.[4] L. L. Veeck,
An atlas of human gametes and conceptuses: an illustratedreference for assisted reproductive technology . CRC Press, 1999.[5] P. Khosravi, E. Kazemi, Q. Zhan, J. E. Malmsten, M. Toschi, P. Zisi-mopoulos, A. Sigaras, S. Lavery, L. A. Cooper, C. Hickman et al. , “Deeplearning enables robust assessment and selection of human blastocystsafter in vitro fertilization,” npj Digital Medicine , vol. 2, no. 1, p. 21,2019.[6] Z. Akkus, A. Galimzianova, A. Hoogi, D. L. Rubin, and B. J. Erickson,“Deep learning for brain mri segmentation: state of the art and futuredirections,”
Journal of digital imaging , vol. 30, no. 4, pp. 449–459, 2017.[7] G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi,M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, and C. I. S´anchez,“A survey on deep learning in medical image analysis,”
Medical imageanalysis , vol. 42, pp. 60–88, 2017.[8] J. A. Quinn, R. Nakasi, P. K. Mugagga, P. Byanyima, W. Lubega,and A. Andama, “Deep convolutional neural networks for microscopy-based point of care diagnostics,” in
Machine Learning for HealthcareConference , 2016, pp. 271–281.[9] N. Zaninovic, O. Elemento, and Z. Rosenwaks, “Artificial intelligence:its applications in reproductive medicine and the assisted reproductivetechnologies,”
Fertility and sterility , vol. 112, no. 1, pp. 28–30, 2019.[10] M. F. Kragh, J. Rimestad, J. Berntsen, and H. Karstoft, “Automaticgrading of human blastocysts from time-lapse imaging,”
Computers inBiology and Medicine , p. 103494, 2019.[11] T.-J. Chen, W.-L. Zheng, C.-H. Liu, I. Huang, H.-H. Lai, and M. Liu,“Using deep learning with large dataset of microscope images to developan automated embryo grading system,”
Fertility & Reproduction , vol. 1,no. 01, pp. 51–56, 2019.[12] S. Coskun, J. Hollanders, S. Al-Hassan, H. Al-Sufyan, H. Al-Mayman,and K. Jaroudi, “Day 5 versus day 3 embryo transfer: a controlledrandomized trial,”
Human Reproduction , vol. 15, no. 9, pp. 1947–1952,2000.[13] S¸. Hatırnaz and M. K. Pektas¸, “Day 3 embryo transfer versus day 5blastocyst transfers: A prospective randomized controlled trial,”
TurkishJournal of Obstetrics and Gynecology , vol. 14, no. 2, p. 82, 2017.[14] K.-C. Lan, F.-J. Huang, Y.-C. Lin, F.-T. Kung, C.-H. Hsieh, H.-W.Huang, P.-H. Tan, and S. Y. Chang, “The predictive value of using acombined z-score and day 3 embryo morphology score in the assessmentof embryo survival on day 5,”
Human reproduction , vol. 18, no. 6, pp.1299–1306, 2003.[15] J. Howard and S. Gugger, “Fastai: A layered api for deep learning,”
Information , vol. 11, no. 2, p. 108, 2020. [16] M. Oquab, L. Bottou, I. Laptev, and J. Sivic, “Learning and transferringmid-level image representations using convolutional neural networks,”in
Proceedings of the IEEE conference on computer vision and patternrecognition , 2014, pp. 1717–1724.[17] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for imagerecognition,” in
Proceedings of the IEEE conference on computer visionand pattern recognition , 2016, pp. 770–778.[18] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Denselyconnected convolutional networks,” in
Proceedings of the IEEE confer-ence on computer vision and pattern recognition , 2017, pp. 4700–4708.[19] F. Chollet, “Xception: Deep learning with depthwise separable convolu-tions,” in
Proceedings of the IEEE conference on computer vision andpattern recognition , 2017, pp. 1251–1258.[20] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen,“Mobilenetv2: Inverted residuals and linear bottlenecks,” in
Proceedingsof the IEEE conference on computer vision and pattern recognition ,2018, pp. 4510–4520.[21] L. N. Smith, “Cyclical learning rates for training neural networks,”in . IEEE, 2017, pp. 464–472.[22] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinkingthe inception architecture for computer vision,” in
Proceedings of theIEEE conference on computer vision and pattern recognition , 2016, pp.2818–2826.[23] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”in
Proceedings of the IEEE conference on computer vision and patternrecognition , 2015, pp. 1–9.[24] R. Pascanu, T. Mikolov, and Y. Bengio, “On the difficulty of trainingrecurrent neural networks,” in
International conference on machinelearning , 2013, pp. 1310–1318.[25] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv preprint arXiv:1804.02767 , 2018.[26] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networksfor biomedical image segmentation,” in