Improving the performance of deep convolutional neural networks (CNN) in embryology using synthetic machine-generated images

Rohan Banerjee
Department of Medicine
Division of Engineering in Medicine
Poster Overview

Even though artificial intelligence-based approaches have shown promising results in the classification of oocyte image, the scarcity of dataset, due to its availability at very few fertility centers, limits its confidence and therefore the performance. Generation of synthetic images leveraging the learning capabilities of artificial intelligence has showed a lot of potential in multiple research applications in the recent times. This tech-stack helps the machines generate images which do not exist in the real world. We propose an unsupervised learning methodology to generate synthetic life-like oocyte images to mitigate the data requirements and in turn increase the performance of the classification. At the end, the generation of the synthetic images and extensive learning cycles, we see a significant increase of 15.58% in the confidence of our proposed classification algorithm.

Scientific Abstract

Objective: CNNs have shown an enormous potential in embryology especially with a large amount of data. However, the availability of datasets is limited to very few fertility centers. Generative adversarial networks (GANs), which can generate life-like images, may help in mitigating such data requirements.

Design: Using a dataset of clinical oocyte images with known fertilization outcomes (KFO) and synthetic oocyte images generated by a pretrained GAN, a CNN, henceforth called synthetic CNN (s-CNN), was trained to classify between oocytes that eventually fertilized normally, with 2 pro-nuclei (2PN),
and abnormally (non-2PN). The network performance was compared to the performance of a CNN, henceforth called the original CNN (o-CNN), trained using clinical oocyte images only.

Method: A GAN was trained to generate life-like oocyte images from scratch. The s-CNN was trained using a dataset of 1411 clinical oocytes images with KFO and 1340 synthetic oocyte images generated by the GAN. The o-CNNs were trained as reported previously using only the 1411 clinical oocyte
images with KFO.

Results: The o-CNN performed with an accuracy of 67.0%. In contrast, the s-CNN performed with an improved accuracy of 82.58% in differentiating the oocyte images based on their eventual fertilization outcomes.

Clinical Implications
The training with GAN-generated data helped s-CNN to outperform the conventionally trained o-CNN when evaluating real oocytes with KFO. Thus, GANs may hold the potential to improve the currently utilized CNNs in embryology.
Research Areas
Manoj Kanakasabapathy, Charlie Bormann, Prudhvi Thirumalaraju, Rohan Banerjee, Hadi Shafiee
Principal Investigator
Hadi Shafiee, PhD

Explore Other Posters

Comments are closed.