Content Subtle Style Generative Adversarial Network with Multitask Efficient NetB7 to Predict Embryo Ploidy Status for Effective IVF Outcomes

Usually, the assisted reproductive units prevent multiple pregnancies by transferring a single embryo. Experts should choose an embryo to transplant from a unit created by the intended parents. This selection technique must be precise, non-intrusive, reliable, and accessible to In-Vitro Fertilization (IVF) facilities globally. However, the key challenge is determining embryo ploidy status and choosing the most viable embryos for successful transfer. To address this issue, embryo genetic status is assessed via Preimplantation Genetic Testing for Aneuploidy (PGT-A), which entails embryo biopsy and genetic testing. However, it is invasive and costly. As a result, non-invasive techniques using Machine Learning (ML) and Deep Learning (DL) models are created to estimate embryo ploidy status from time-lapse embryo photos, allowing for decision-making before further treatment. Conversely, these models frequently rely on low-quality and insufficient numbers of embryo images as well as a lack of clinical and demographic information about the patient, such as age, pelvic factor, sperm count, etc., resulting in decreased accuracy. For augmenting embryo images, Generative Adversarial Network (GAN) and its variant called Style-based GAN (StyleGAN) were applied; yet, they were restricted to producing images with low resolutions, leading to poor image quality. Hence, this article develops a novel Content Subtle Style-based GAN (CSSGAN) with a Multi-model EfficientNetB7 (MENB7) model to augment time-lapse embryo images and predict ploidy status for effective IVF outcomes. The CSSGAN adopts content-aware channel pruning and content-sensitive distillation approaches as minimax optimization tasks to generate a massive quantity of high-quality and high-fidelity human embryo images at the blastocyst stage. The generated images are input into the Mask-Residual Convolutional Neural Network (Mask-RCNN) for segmentation. Then, the relevant features from the segmented images are captured by EfficientNetB7, which fuses them with the patient’s clinical and demographic data to create a unified feature vector. Moreover, a Fully Connected (FC) layer with softmax is used for predicting euploid and aneuploid embryos. Finally, extensive experiments using distinct datasets illustrate the efficiency of the CSSGAN-MENB7 model in predicting embryo ploidy status with 95.88% accuracy compared to the conventional DL models.