Ensemble deep learning with advanced feature engineering for embryo evaluation on in-vitro fertilisation procedures using biomedical images

In vitro fertilisation (IVF) is the most commonly used assisted reproductive technology employed to treat infertility. It includes intricate treatment processes such as egg retrieval, fertilisation, ovulation induction, embryo transfer, and implantation. The lower success rate of IVF was projected owing to the poor quality of embryos; thus, many embryos are frequently moved, raising difficulties for children and mothers, as well as increasing the cost of healthcare. A precise assessment of embryo grade, size, and developmental stage is significant in an embryo transfer program. Completely automatic embryo assessments, in which quality grade is assigned without user intervention from an image-analysis perspective, are complex owing to the intricacy of embryo morphology. Conversely, developments in deep learning (DL) have enabled precise, intention-based classification of images across non-medical and medical domains without the need for labour-intensive feature engineering. This article proposes an ensemble deep learning-enabled embryo evaluation system using advanced feature engineering of biomedical images (EDLEVS-AFEBI) model in IVF procedures. This paper's primary aim is to propose an automated embryo grading method using advanced techniques to improve selection for successful implantation and pregnancy outcomes. At first, the image pre-processing phase uses an adaptive Gaussian bilateral filter (AGBF) to enhance image quality by removing noise. The improved DenseNet model is employed for the feature extraction process to recognise and isolate the most relevant information from raw data. Finally, ensemble learning models such as the temporal convolutional network (TCN), the Elman neural network (ENN), and the conditional variational autoencoder (CVAE) are utilised for embryo classification. The comparison analysis of the EDLEVS-AFEBI methodology showed an accuracy of 99.39% compared with other models on the microscopic image dataset.