Table 4.
Comparison table of studies associated with the embryo grading task
| Paper | AI Architecture | Input data type | Training data size | CNN function | Function limit | Prediction accuracy |
|---|---|---|---|---|---|---|
| Kragh et al. (2019) [58] | Xception CNN model [57] paired with custom RNN model | EmbryoScope time-lapse images | 8.664 embryo | Assigning ICM and TE grades | – | ICM ± 65%, TE ± 69% |
| Khosravi et al. (2019) [60] | Inception-V1 model [41] along with CHAID decision tree [64] | EmbryoScope time-lapse image | 12,001 time-lapse image up to seven focal depths | Classifying embryo into good and poor quality & predicting pregnancy probability | – | Grading precision: 95.7% |
| Chen et al. (2019) [59] | ResNet50 model [40] | Microscopic images | 171,239 static images from 16,201 embryos of 4,146 IVF cycles | Automatic grading based on Blastocysts Development, ICM, and TE grades | Only detected D5/D6 embryo images for prediction | Blastocyst: 96.24%, TE: 84.42%, ICM: 91.08% |
| VerMilyea et al. (2020) [54] | Ensemble of multiple ResNet models [40] and DenseNet modesl [31] | Optical light microscope images | 1667 microscopic images of embryo in Day 5 | Predicting embryo viability | Analysis only on Day 5 embryos | 64.3% prediction accuracy |
| Bormann et al. (2020) [61] | Xception [57] with Genetic Algorithm | Vitro-life Embryoscope time-lapse videos | 2440 static human embryo images recorded at 113hpi | Embryo selection based on embryo quality and implantation outcome | Only utilizes images from single timepoint even with time-lapse | Based on: embryo quality = 90.97% implantation outcome = 82.76% |
| Kanakasabapathy et al. (2020) [55] | Xception [57] + Genetic Algorithm | 3469 recorded video from 543 patients | Training: 1190, Validation: 511 | Embryo selection based on morphology quality | – | SET: ± 83.51% DET: ± 96.90% |
| Thirumalaraju et al. (2020) [62] | Multiple tested, Xception [57] deemed as best | Embryoscope time-lapse system | Training: 1188 images Validation: 510 images Additional 742 as test | Grading blastocysts | – | ± 90.90% |
| Silver et al. (2020) [63] | An implementation of CNN (no specific architecture were named) | Time-lapse videos | 8789 videos in total Additionally 272 were KID labeled | Embryo grading and implantation prediction | – | AUC score of ± 0.82% |