Segmentation of mature human oocytes provides interpretable and improved blastocyst outcome predictions by a machine learning model

Abstract

Within the medical field of human assisted reproductive technology, a method for interpretable, non-invasive, and objective oocyte evaluation is lacking. To address this clinical gap, a workflow utilizing machine learning techniques has been developed involving automatic multi-class segmentation of two-dimensional images, morphometric analysis, and prediction of developmental outcomes of mature denuded oocytes based on feature extraction and clinical variables. Two separate models have been developed for this purpose-a model to perform multiclass segmentation, and a classifier model to classify oocytes as likely or unlikely to develop into a blastocyst (Day 5-7 embryo). The segmentation model is highly accurate at segmenting the oocyte, ensuring high-quality segmented images (masks) are utilized as inputs for the classifier model (mask model). The mask model displayed an area under the curve (AUC) of 0.63, a sensitivity of 0.51, and a specificity of 0.66 on the test set. The AUC underwent a reduction to 0.57 when features extracted from the ooplasm were removed, suggesting the ooplasm holds the information most pertinent to oocyte developmental competence. The mask model was further compared to a deep learning model, which also utilized the segmented images as inputs. The performance of both models combined in an ensemble model was evaluated, showing an improvement (AUC 0.67) compared to either model alone. The results of this study indicate that direct assessments of the oocyte are warranted, providing the first objective insights into key features for developmental competence, a step above the current standard of care-solely utilizing oocyte age as a proxy for quality.

Figure 1

Segmentation of the ooplasm, perivitelline space, and zona pellucida. (a) Ground truth labels assigned by embryologists. (b) Unsegmented oocyte. (c) Segmentation of ooplasm. (d) Segmentation of perivitelline space. (e) Segmentation of zona pellucida.

Figure 2

Mean Shapley values, ranking features used by the LightGBM mask model by importance to model predictions. (A) Mean Shapley values across entire model development dataset. (B) Mean Shapley values across external validation dataset.

Figure 3

Waterfall plots visually demonstrating how model predictions are made. (A) Negative prediction. Starting from a prior expectation of − 0.221, the Shapley values of the features used by the model are added up to generate a value f(x) =  − 0.654—the logit of the model output that is then inputted into the sigmoid function to generate a prediction probability. (B) Positive prediction. Starting from a prior expectation of − 0.221, the Shapley values of the features are added up to generate the logit of the model output, f(x) = 0.17.

Figure 4

Subgroup analysis by patient age group for the mask and ensemble model. (A) Subgroup analysis by age group for the mask model displayed significant difference in performance only for the 38–39 age group (AUC 0.6) compared to model performance on the entire dataset (AUC 0.63) (p < 0.05, DeLong test). (B) Subgroup analysis by age group for the ensemble model displayed no significant differences in performance between the age groups and the entire dataset.

Figure 5

Subgroup analysis by clinic location for the mask and ensemble model. (A) Subgroup analysis by clinic location for the mask model displayed significantly higher performance for the Spain 2 clinic (AUC 0.71, p < 0.01 DeLong test) and the Czechia clinic (AUC 0.75, p < 0.01 DeLong test) compared to performance on the entire dataset (AUC 0.63). Performance on the Canada clinic displayed significant difference (AUC 0.62, p = 0.0496 DeLong test), however, with a borderline effect. B) Subgroup analysis by clinic location for the ensemble model displayed significant differences in performance for the Spain 2 (AUC 0.72, p < 0.01 DeLong test), Czechia (AUC 0.79, p < 0.01 DeLong test), and India (AUC 0.57, p < 0.01 DeLong test) clinic locations compared to the entire dataset (AUC 0.67).

Figure 6

The proposed workflow utilizes two models—the first creates masks for each image of the oocyte, which is then used along with clinical variables as inputs into a classifier model (mask model) to generate a prediction of blastocyst development.