Computer Aided Grading System of Digital Microscopic Blastocyst Images

Abstract

Blastocyst grading is among the critical factors that influence the success of in vitro fertilization (IVF) treatment cycles. Blastocyst morphology grading is traditionally performed through manual microscope examinations. Manual microscopic blastocyst morphological grading is a time-consuming task that suffers from intraobserver and interobserver variation. Therefore, automation of blastocyst grading is essential for IVF success. In this paper, we propose a computer-aided grading system for blastocyst images based on Gardner's grading system. Gardner's grading system consists of three components that correspond to specific regions of the blastocyst. Each component has its own classes. The first component, Expansion, is graded into six grades ranging from 1 to 6. The second component is inner cell mass (ICM) grading into three grades (A-B-C). The third component is trophectoderm (TE) grading into three grades (A-B-C). The proposed system is comprised of three basic stages: dataset acquisition, data preparation, and classification. The dataset was acquired from the “Boy and Girl” clinic center, Cairo, Egypt. The dataset comprises 1015 blastocyst images, extracted from 651 images captured by inverted microscope “Nikon eclipse Ti-U” with a resolution of 640 × 480 pixels. The data preparation stage comprises of blastocysts extraction and localization followed by blastocyst labeling by an experienced embryologist. Data augmentation was, later, performed to enhance the robustness and generalize the capability of the trained models on limited datasets. Subsequently, this work contributes by employing many convolutional neural networks (CNNs) including: VGG16, RESNET50, MobileNetv2, EfficientNetB0, and YOLOv8 to choose the best classification framework for blastocysts. The novelty of our work is based on full automation of standard Gardner's grading system using single static microscopic image. The results showed that the fine YOLOv8 framework achieved the highest accuracy of 97%, 82%, and 89% for expansion, TE, and ICM, respectively.