A Novel Integrative Non-invasive Embryo Selection Approach for in Vitro Fertilization Based on Artificial Intelligence Enhanced Morphokinetic Analysis and Raman Spectra in Spent Culture Media

Our study will include two phases. In Phase I (a retrospective study), archived data will be collected to train the CNN-enhanced MK with RS method on embryo selection, leading to the integrated approach (MK-RS). In Phase II (a prospective study), the integrated MK-RS method established will be used to select embryos, assess the clinical pregnancy rate and evaluate the efficacy of our approach in a randomized controlled trial.

In Phase I, images of the embryo will be captured every 10 minutes by the in-built microscope and camera in the time-lapse incubator. Images will then be assessed by the CNN algorithm for day one human embryo segmentation to identify three distinct features: the zona pellucida (ZP), cytoplasm and pronucleus (PN). The morphodynamics of these three features during the fertilisation to first division will be wrapped up as time series data for the integration. The morphology changes after the first division will be semi-auto annotated, which will be analysed by the commercial MK scoring system (KID Score).

After removal of embryos/blastocysts from the culture dish, the corresponding spent culture medium (SCM) will be collected in sterile polymerase chain reaction (PCR) tubes. Blank culture mediums will also be collected with the same operating standard. A specifically designed sampler will be used to pipette 7μL SCM of each sample, passing through the oil layer of the SCM, and then drop it onto a disposable quartz glass slide and illuminated by RS system (Basecare Raman 200, China). The RS system will be calibrated to 520.5 cm-1 by silicon wafer before testing. Laser excitation parameters are set as follows: 785 nm wavelength, 320 mW power and 100 μm laser spot diameter. Signals are captured in standard mode with a chargecoupled device (CCD) camera with a 20-seconds integration time. Three replicates will be done for each aliquot. Re-calibration is essential when different culture media is tested, considering that G-1 medium is used for embryos before Day 3 and G-2 medium is used for embryos after Day 3. All obtained spectra will be pre-processed by subtracting the background signal. Spectroscopy signals within the near-infrared region (600 cm-1- 1800 cm-1) are analysed for vector normalisation using Labspec 6 software (Horiba, Japan). Our previous SCM samples with known TE ploidy results will be used as a training dataset to establish euploid-aneuploid classification standards. Stacking classification algorithm will be adopted, considering its high overall accuracy (95.9%, unpublished data).

As the segmented time series of CNN algorithm, KID Score annotations and RS profiling results have hundreds of subparameters, we will assemble them together with the ensemble learning, which considers each subparameter as a weak classifier and re-allocated their weights during the training. The primary index for the training is the clinical pregnancy outcome and for the ultimation of the information beneath CNN-enhanced MK and RS, the blastocyst formation results will be used as a secondary index. The ratio of the training set and test set will be 1:1 and in the training set, a 5-fold cross-validation will be performed for monitoring the overfitting.

Phase II will comprise a prospective, single-blinded, randomised controlled trial designed to validate the trained MK-RS method for embryo selection. Metabolomic SCM profiling using RS with CNN-enhanced MK analysis will be used to assess embryo developmental potential along with traditional morphological embryo assessments. The embryo developmental potential results will be used to select the best quality of embryos. Sensitivity and specificity will be assessed using the patient's TE biopsy or non-invasive prenatal testing (NIPT) results will further confirm the scoring of MK-RS method, if available.