Robotic Manipulation of Cumulus–Oocyte Complexes for Cumulus Cell Removal

Abstract

The removal of cumulus cells from cumulus–oocyte complexes is a critical step in clinical in vitro fertilization. Since the oocyte is partially occluded by the surrounding cumulus cells and individual cumulus cells are small in size, it is difficult for embryologists to assess the oocyte's maturity before cumulus cell removal and to completely remove all the cumulus cells manually . Furthermore, it is easy for the oocyte to become lost inside the micropipette during aspiration due to the inaccuracy of manual control. To deal with these difficulties, a robotic system was developed to completely remove cumulus cells from mature oocytes. In this study, an EPSANet50 network was developed to accurately assess the maturity of oocytes, avoiding the removal of cumulus cells around the immature oocyte. An adaptive controller was designed to accurately position oocytes at the target position, reducing the loss of oocytes inside the micropipette. An improved Yolov5s network was proposed to quantify the number and size of cumulus cells and assess the completeness of cumulus cell removal. The experimental results on mouse cumulus–oocyte complexes showed that the robotic system had a higher success rate (98.0 ± 1.8% vs. 85.3 ± 2.4%) and lower discard rate (4.1 ± 2.7% vs. 19.6 ± 3.5%) than the manual operation. Moreover, a higher amplification rate and lower non-specific rate were also achieved by the robotic system in the subsequent genetic testing procedure, indicating reduced genetic contamination from the cumulus cells.

1. Introduction

The removal of cumulus cells from cumulus–oocyte complexes (COCs) is a critical step before intracytoplasmic sperm injection (ICSI), which is an infertility treatment used in around 70% of in vitro fertilization (IVF) cycles [1]. A COC is composed of an oocyte and the surrounding cumulus cells, which promote oocyte maturation [2]. However, cumulus cells block the injection micropipette during oocyte microinjection in ICSI. Moreover, the presence of cumulus cells also causes genetic contamination in the subsequent preimplantation genetic testing (PGT) due to the mixture of the target embryo and the non-target cumulus cells in the test sample [3,4].

In manual cumulus cell removal, the COCs retrieved from ovarian follicles are first mixed with hyaluronidase to loosen the cumulus–oocyte bonds. The COCs are then transferred to M2 culture medium, and an embryologist manually removes cumulus cells by repeatedly aspirating and dispensing the COCs using a micropipette [5], as shown in Figure 1a. After cumulus cell removal, the mature oocytes with visible polar bodies are fertilized via ICSI and further cultured to blastocysts (i.e., day 5/6 embryos). Several trophectoderm (TE) cells are collected from the blastocysts for PGT, and blastocysts without genetic abnormalities are selected for the subsequent embryo transfer. In manual operations, embryologists remove cumulus cells without assessing the oocyte's maturity due to the occlusion of the polar body by the surrounding cumulus cells. The removal of cumulus cells before oocyte maturation may affect the maturation process and eventually result in maturation failure [6]. During cumulus cell removal, it is difficult to position the COC inside the micropipette, since the mass of the COC varies during aspiration. The COCs tend to move deeper into the micropipette and disappear, leading to the loss of COCs. Furthermore, since cumulus cells are small in size, it is difficult for embryologists to determine whether the cumulus cells have been completely removed from the oocyte. The remaining cumulus cells on the oocyte will lead to ICSI failure by blocking the injection micropipette, and they also reduce the PGT accuracy by introducing genetic materials different from the tested embryos, as shown in Figure 1b,c.

There have been several emerging trends in robotics and automation in recent years. Diachenko et al. integrated 3D digital twins into a collaborative robot, enabling robot joint visualization and automated control [7]. Walker et al. proposed a virtual surrogate robot based on augmented reality technologies to automatically foreshadow the actions of a physical robot and improve the robot teleoperations [8]. Cecil et al. proposed a cyber–physical framework, which was capable of accomplishing automated assembly for micro-devices [9]. Nadrag et al. used the zoom effect to deal with delay masking in order to remotely automate a robot through the Internet [10]. Hörner et al. developed a holographic optical tweezers-based platform to automatically manipulate organelles and characterize the mechanobiological properties of cells [11]. Robotic techniques are emerging in the areas of virtual robotics and physical robotics.

Techniques have been developed to automatically remove cumulus cells. Han et al. designed a microfluidic device consisting of a digesting chamber and a pit array, where COCs go through a long channel filled with hyaluronidase for cumulus cell removal and then fall into pits due to gravity for fertilization [12]. Nguyen et al. proposed an active microfluidic device with multiple inlet ports to remove cumulus cells. By using multiple inlet ports, the device avoided clogging COCs during cumulus cell removal [13]. Min and Ma developed a vortex device to remove cumulus cells by producing a vortex around the oocytes [14]. We previously proposed a robotic system that removed cumulus cells by repeatedly aspirating and dispensing the COCs via motion control of the COCs inside the micropipette [15]. However, these devices focused on the pressure control of the flow inside the microfluidic channels or micropipettes. The assessment of oocyte maturity before cumulus cell removal and the quantification of the completeness of cumulus cell removal were ignored. Furthermore, the mass of the COCs during cumulus cell removal was considered to be constant, resulting in large positioning errors of the COCs inside the channels or micropipettes.

Several methods have been developed to assess oocyte maturity. Since the oocyte is considered mature when the polar body is extruded from the oocyte and clearly visible [16], Wang et al. introduced an edge detection method to identify the presence and orientation of the polar body of the oocyte that is rotated by the injection micropipette [17]. Chen et al. proposed an improved support vector machine method for the detection of the polar body in order to aid in the removal of the polar bodies from the oocyte [18]. Dai et al. used the deep neural network method to robustly detect polar bodies with various sizes and shapes for orientation control of the oocyte [19]. However, the existing methods only detect polar bodies of oocytes after cumulus cell removal. In this scenario, the cumulus cells surrounding the oocytes will degrade the detection performance of the current methods.

To accurately position the cells at the target position inside the micropipette, Shakoor et al. developed a sliding controller to position the fibroblast cell with less than 20 μm in diameter at the target position within the microfluidic channel during organelle biopsy, enabling the 3D confinement of small organelles for extraction [20]. Zhu et al. designed a PID controller to decelerate oocytes ~100 µm in diameter to the target position inside the micropipette during vitrification, preventing the oocyte from becoming lost out of the field of view [21]. However, the existing methods only consider cells with constant mass. During aspiration and deposition, the mass of COC varies as the cumulus cells are removed. Ignoring the mass change in the COC will lead to the large positioning errors.

To assess whether cumulus cells have been completely removed, machine learning models for cumulus cell detection have been proposed. Baručić et al. used a U-shaped network model to detect cumulus cells and classify the removal completeness [22]. Sánchez et al. introduced the model based on a faster region-based convolutional neural network to assess the removal completeness [23]. However, the existing methods focus on the qualitative assessment of removal completeness. Quantitative methods are needed to accurately calculate the number of remaining cumulus cells to ensure complete removal.

In this work, a physical robotic system was developed to automatically remove cumulus cells. The challenges to overcome are as follows. (1) It is difficult to assess the maturity of an oocyte via the polar body detection, as the polar body is occluded by the surrounding cumulus cells. (2) The mass of the COC varies during micropipette aspiration and dispensing, leading to the large positioning errors of the COC inside the micropipette. Therefore, it is easy for the COC to become lost after it enters deep into the micropipette using existing aspiration control. (3) It is difficult to accurately calculate the number of the remaining cumulus cells surrounding the oocyte, as the cumulus cells are small, varied in size, and often appear blurred due to out-of-focus imaging conditions. To tackle these challenges, we developed a new physical robotic system with AI-powered technologies. An EPSANet50 network was developed to accurately assess the maturity of oocytes based on the features of the polar body. An adaptive controller was designed to compensate for the positioning errors caused by the varying COC mass and to accurately position the COC at the target position inside the micropipette. An improved Yolov5s network was proposed to accurately detect cumulus cells for the quantitative assessment. The robotic system aims to automate and standardize the removal of cumulus cells from COCs in IVF. Compared with manual operation, the robotic system showed a higher success rate for complete cumulus cell removal and a lower discard rate of immature oocytes. Furthermore, the experimental results demonstrated that the robotic system achieved a higher amplification rate and a lower non-specific rate for genetic testing compared with manual operation, which had the potential to significantly improve clinical practices in IVF.

2. System Overview

Figure 2a shows the system setup for the automated removal of cumulus cells. A two-axis X–Y stage (HLD117, Prior, travel range: 120 mm × 72 mm) equipped with a motorized focusing module (FM-100, Suzhou Boundless Medical Technology Co., Ltd., Suzhou, China, resolution: 0.1 mm/r) was used to position COCs in the field of view (FOV) within the focal plane. A micropipette (WG-DD-125, WEGO, Singapore, inner diameter: 125 μm) was mounted on a four-axis micromanipulator (MX7600, Siskiyou, Grants Pass, OR, USA, resolution: 0.1 μm) for manipulating COCs. A motorized oil pump (CellTram 4r Oil, Eppendorf, resolution: 1 μL/r) was connected to the micropipette to control the pressure at the micropipette tip for aspirating and dispensing COCs. A camera (acA1300-60gm, Basler, Ahrensburg, Germany, frame rate: 60 frames/s) was connected to the microscope captures visual feedback via bright-field imaging. Visual feedback of the micropipette and COCs was used for position control and to assess both oocyte’s maturity and the completeness of cumulus cell removal.

Female C57 mice (8–12 weeks old, Nanjing Aibei Biotechnology Co., Ltd., Nanjing, China) were injected with PMSG and HCG hormones, and each ampulla of the oviduct was incised to release COCs after injection. The COCs were then enzymatically treated with hyaluronidase and transferred to a droplet of M2 medium for cumulus cell removal. During the automated cumulus cell removal process, the X-Y stage was controlled to position the target COC in the FOV, and the focal plane was adjusted to the middle plane of the target COC using the motorized focusing module. To avoid manipulating immature oocytes, the maturity of the oocyte was evaluated by an EPSANet50 network before removing the cumulus cells from the COC. Then, the micropipette tip was automatically positioned 140 μm to the left of the COC within the focal plane. The COC was repeatedly aspirated into the micropipette, positioned at the target position, and then dispensed outside the micropipette by an adaptive controller. After COC dispensing, the remaining cumulus cells around the oocyte were detected by an improved Yolov5s network to quantitatively assess the completeness of cumulus cell removal. The oocyte was then transferred to a droplet of KSOM medium for subsequent ICSI and PGT procedures. The workflow of the robotic system is described in Figure 2b.

3. Methods

3.1. Assessment of Oocyte Maturity

In clinical operation, embryologists assess an oocyte’s maturity by observing the presence of the polar body after the removal of cumulus cells. If the polar body is not observed, the oocyte will be further cultured for another 24 hours. Immature oocytes after further culturing are not qualified for ICSI and will be discarded. Recent studies have shown that communication between cumulus cells and the oocyte is essential for oocyte growth and maturation [24]. Therefore, removing cumulus cells before an oocyte is fully mature can affect its maturation process and eventually lead to maturation failure.

An EPSANet50 network [25] is proposed to assess the maturity of the oocyte from the COC image, avoiding cumulus cell removal from an immature oocyte. As shown in Figure 3a, the proposed EPSANet50 network consists of 16 EPSANet blocks and one output module. In each EPSANet block, a pyramid squeeze attention (PSA) module [26] (see Figure 3b) was utilized to extract multi-scale features and the weighted channel attention, which can help improve the extraction ability of COC features and focus on key features such as the polar body. The high-level feature map F_i obtained from the i-th EPSANet block is

$$F_i=Con{v}_{\left(1\times 1,1\right)}\left(Y_i\right)=Con{v}_{\left(1\times 1,1\right)}\left(S_i\otimes Softmax\left(Seweight\left(S_i\right)\right)\right),i=1,2\cdots ,16$$

(1)

where Conv is the convolution operation, Y_i is the output feature map of the PSA module, ⊗ is the multiplication in the channel dimension, Softmax is the activation function for the output part, Seweight is the Seweight module used to extract the weighted channel attention, and the input feature map S_i of the Seweight module is

$$S_i=Con{v}_{\left(3\times 3,2\right)}\left(X_i\right)\oplus Con{v}_{\left(5\times 5,4\right)}\left(X_i\right)$$

(2)

where ⊕ is the splicing in channel dimension, and X_i is the input feature map of the EPSANet block. Substituting Equation (2) into Equation (1), the high-level feature map F_i of the COC is

$$F_i=Con{v}_{\left(1\times 1,1\right)}\left(PSA\left(X_i\right)\right),i=1,2\cdots ,16$$

(3)

where

$$PSA\left(X_i\right)=\left(Con{v}_{\left(3\times 3,2\right)}\left(X_i\right)\oplus Con{v}_{\left(5\times 5,4\right)}\left(X_i\right)\right)\otimes Softmax\left(Seweight\left(Con{v}_{\left(3\times 3,2\right)}\left(X_i\right)\oplus Con{v}_{\left(5\times 5,4\right)}\left(X_i\right)\right)\right)$$

In the output module, the mean pixel value of the highest-level feature map F_i (where i = 16) is calculated by the average pooling operation Average(·) followed by the fully connected operation F_c. Then, the Softmax activation function Softmax(·) is used to predict the probability C of the oocyte maturity as

$$C=Softmax\left(F_c\left(Average\left(F_{i=16}\right)\right)\right)=Softmax\left(F_c\left(Average\left(Con{v}_{\left(1\times 1,1\right)}\left(PSA\left(X_{i=16}\right)\right)\right)\right)\right)$$

(4)

3.2. Design of Controller

After assessing the oocyte’s maturity, the COC is repeatedly aspirated and dispensed by the micropipette to remove the cumulus cells. During the micropipette aspiration, the COC is required to be accurately positioned at the target position, aiming to avoid oocyte loss inside the micropipette. However, as the cumulus cells are removed during aspiration and dispensing, the mass of the COC changes. If a fixed external force is applied to it, the acceleration of the COC will change, leading to the large positioning errors. Therefore, the mass change in the COC cannot be ignored and should be updated in real time to compensate for the positioning errors based on the dynamics of the COC’s motion inside the micropipette.

The dynamics of the COC’s motion shown in Figure 4 are described as

$$M\ddot{x}=F_d=6\pi \mu r_o\left(\gamma {\upsilon }_p-\dot{x}\right)$$

(5)

where M and x represent the mass and position of the COC inside the micropipette, respectively. F_d is the drag force [27], μ is the viscosity of the medium, r_o is the radius of the oocyte, γ is the ratio of the cross-sectional areas of the pump piston and the micropipette, and υ_p is the velocity of the pump piston. The M is

$$M=\frac{4}{3}\pi \left({r_o}^3+{\displaystyle \sum _{i=1}^N{r_i}^3}\right)\rho$$

(6)

where N is the number of cumulus cells, r_i is the radius of the i-th cumulus cell, and ρ is the density of the cell.

Substituting Equation (6) into Equation (5) obtains

$$(\beta {r_o}^2\ddot{x}+\frac{1}{\gamma }\dot{x})+\frac{\beta {\displaystyle \sum _{i=1}^N{r_i}^3}}{r_o}\ddot{x}={\upsilon }_p$$

(7)

where

$$\beta =\frac{2\rho }{9\mu \gamma }$$

In Equation (7), β, r_o, and γ are constants, but N varies during the micropipette aspiration, leading to changes in the total mass of cumulus cells.

To address the varying mass of cumulus cells, an adaptive controller has been proposed to control the piston velocity for accurate COC positioning inside the micropipette. In the adaptive control law, u₁ and u₂ represent the feedback and compensate terms, respectively.

$$u_1=k_{p}e+k_d\dot{e}, u_2=\frac{{\displaystyle \sum _{i=1}^N{r_i}^3}}{{r_o}^3}(k_{p}e+k_d\dot{e}-\frac{1}{\gamma }\dot{x})$$

(8)

where k_p and k_d are the proportional and derivative gains, respectively, e = x − x_d is the positioning error, and x_d is the target position. u₂ is used to update N and the corresponding r_i, aiming to compensate for the positioning errors caused by the varying mass of cumulus cells. N and r_i can be estimated by the network for cumulus cell detection in Section 3.3.

Finding the control law u = u₁ + u₂ is equivalent to minimizing the cost function.

$$J={\displaystyle {\int }_{t_0}^{t_f}(X^{T}QX+{u_1}^{T}R{u}_1)}dt$$

(9)

where t₀ and t_f are the time for the beginning and end, respectively; Q and R are the weights for the state and control, respectively; u₁^TRu₁ and X^TQX represent the input energy and position error, respectively. The minimum of Equation (9) is solved based on the linear quadratic regulator (LQR) theory, and the u₁ is

$$u_1=-Kx=-R^{-1}BPx$$

(10)

where K = [k_p, k_d] is the feedback matrix, and A and B are coefficient matrices. Based on the Riccati equation, P is

$$A^{T}P+PA-PB{R}^{-1}{B}^{T}P+Q=0$$

(11)

As shown in Figure 5, the adaptive control, including an LQR and adaptive compensation, is proposed to adjust the piston velocity based on the positioning errors of the COC. The LQR is used to control the main body of the COC (i.e., oocyte), and the adaptive compensation is used to further adjust the piston velocity based on the mass changes of cumulus cells. During the aspiration, the position x of the COC, which serves as the control output, is obtained from visual tracking. The Kalman filter [28] is used to overcome the interference from impurities inside the micropipette, ensuring tracking accuracy.

3.3. Detection of Cumulus Cells

Since cumulus cells are small and appear blurred due to out-of-focus imaging, it is difficult to determine the number and radius of cumulus cells within a COC. Previous studies used binary classification networks that qualitatively assess the cumulus cells surrounding the oocyte. However, these networks are unable to accurately detect the number of cumulus cells and show relatively large classification errors when a few small and blurry cumulus cells remain on the oocyte [22,23].

In this study, an improved Yolov5s network is proposed to accurately detect cumulus cells for the quantitative assessment. As shown in Figure 6, the improved Yolov5s network consists of three major parts: the backbone, neck, and head. The features of the COC are first extracted by the backbone. Then, the extracted features are further fused by the neck. Finally, the head processes the result of feature fusion and outputs the detected cumulus cells with rectangular bounding boxes. The cumulus cell removal rate r is defined as

$$r=\frac{R_{Before}-R_{After}}{R_{Before}}$$

(12)

where R_Before is the number of cumulus cells detected before removal, and R_After is the number detected after removal. r ranges from 0 to 1, and r = 1 means the complete removal of cumulus cells. Compared with the traditional Yolov5s [29], several improvements have been made in the proposed Yolov5s to increase the accuracy of cumulus cell detection as follows.

In the backbone, the traditional Yolov5s uses a CSPDarknet53 module [30] to repeatedly calculate the gradient between two connecting neuronal nodes for backpropagation, which reduces the network’s capability of feature expressiveness and extraction. To overcome this limitation, the CSPDarknet53 module is replaced by the Res2Net-COT module [31] in the improved Yolov5s, as shown in Figure 7a. The input COC feature map is divided into several sub-feature maps X_i, which is followed by feature transformation K_i and feature fusion Y_i. The feature fusion Y_i is defined as

$$Y_i=\left\{\begin{array}{c}{X}_i,i=1\\ K_i\left(X_i+Y_{i-1}\right),1<i<s\end{array}\right.$$

(13)

where s is the number of sub-feature maps. This structure preserves multi-scale feature information and increases the receptive field, which helps alleviate the problem of gradient repetition and enhances the semantic expressiveness of cumulus cell features. Additionally, the COT block embedded within the Res2Net-COT module captures richer contextual information and extracts more useful cumulus cell features, leading to improved detection accuracy.

In the neck, the FPN-PAN module [32] of the traditional Yolov5s has several limitations, such as lower feature utilization and excessive redundancy, which lead to insufficient feature fusion and incomplete feature transmission. To overcome these issues, the BiFPN module [33] is introduced to replace the FPN-PAN module, as shown in Figure 7b. The BiFPN module constructs a bidirectional feature pyramid structure by employing both top–down and bottom–up paths, enhancing the fusion of cumulus cell features at different scales and improving the feature transmission capability.

In the head, the CBS module [34] used in the traditional Yolov5s adopts a serial transmission structure. This structure causes the feature maps of the three detection branches to have high latitude characteristics, increasing the model computation and thus reducing the inference speed of the model. In the improved Yolov5s, the CBS module with a serial transmission structure is replaced by the inception module [35] with a parallel transmission structure, as shown in Figure 7c. The COC feature map M_i is first input into a parallel structure with two branches. One branch consists of a 1 × 1 convolutional layer followed by 1 × 3 and 3 × 1 convolution layers, and the other branch contains only a 1 × 1 convolutional layer. Then, two branches are spliced to restore the number of feature channels. Through the parallel transmission structure, the efficiency of feature transformation is improved, and the amount of calculation is reduced. The output feature map N_i of the inception module is

$$N_i=Concat\left(Con{v}_{(1\times 3)}\left(Con{v}_{(1\times 1)}\left(M_i\right)\right),Con{v}_{(3\times 1)}\left(Con{v}_{(1\times 1)}\left(M_i\right)\right),Con{v}_{(1\times 1)}\left(M_i\right)\right),i=1,2,3$$

(14)

where Concat is the splicing in the channel dimension, and Conv is the convolution operation.

4. Experimental Results

4.1. Oocyte Maturity

To train the EPSANet for oocyte maturity assessment, 1800 images were captured from 180 mouse COCs. For each COC, an image was captured every 12 seconds after enzymatic treatment with hyaluronidase for 2 min, mimicking different concentrations of cumulus cells surrounding the oocyte. The dataset was divided into 80% for training , 10% for validation, and 10% for testing. After image capture, the experienced embryologists manually removed cumulus cells and assessed the maturity of oocytes by observing the presence of the polar body. The assessment results (i.e., mature or immature) were used as the ground truth labels. The performance of the EPSANet50 network was compared with that of the ResNet50 [36], the DenseNet50 [37], and the SENet50 [38] networks, including accuracy, F1-score, and AUC. The accuracy and F1-score are defined as

$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN},F1-score=\frac{2TP}{2TP+FP+FN}$$

(15)

where TP is the number of mature oocytes correctly classified as mature, TN is the number of immature oocytes correctly classified as immature, FP is the number of immature oocytes incorrectly classified as mature, and FN is the number of mature oocytes incorrectly classified as immature. The area under the receiver operating characteristic (ROC) curve (AUC) is calculated to evaluate the model accuracy regardless of the classification threshold.

As shown in Figure 8, the EPSANet50 network achieved an accuracy of 94.5%, an F1-score of 91.2%, and an AUC of 0.91, all of which outperformed the other three networks. The performance of the DenseNet50 network was similar to that of the SENet50 network in terms of accuracy (92.0% vs. 92.3%), F1-score (88.9% vs. 89.2%), and AUC (0.85 vs. 0.86), respectively. In comparison, the ResNet50 network had lower performance, i.e., 89.7% in accuracy, 85.6% in F1-score, and 0.78 in AUC.

4.2. Control Performance

To validate the control performance of the adaptive controller, the COCs were aspirated into the micropipette. The micropipette tip was considered the initial position, and the target position was set at 350 μm from the initial position. Under the same experimental conditions, COCs were aspirated into the micropipette using three controllers, including the adaptive controller, LQR controller, and PD controller. As shown in Figure 9, compared with the LQR and PD controllers, the adaptive controller achieved a shorter settling time (1.58 s vs. 1.86 s vs. 2.23 s). For a total of 270 COCs, the adaptive controller also achieved a smaller overshoot (9.05 ± 0.13 μm vs. 18.17 ± 0.16 μm vs. 29.32 μm ± 0.18 μm), and a smaller positioning error (0.56 ± 0.11 μm vs. 2.73 ± 0.12 μm vs. 6.02 ± 0.16 μm), outperforming the LQR and PD controllers.

4.3. Cumulus Cell Detection

During the cumulus cell removal process, COC images were captured outside the micropipette and resized to 224 × 224 as the input of the improved Yolov5s. A total of 2200 COC images were captured, with 1320 used for training, 440 for validation, and 440 for testing. The COC images were zoomed in, and cumulus cells were labeled with rectangular bounding boxes by the experienced embryologists using the LabelImg software (version: 1.8.1). Both the improved Yolov5s and the traditional Yolov5s were trained on the same dataset, and the validation loss curves are shown in Figure 10. As the number of epochs increased, the loss values decreased rapidly within the first 60 epochs and finally converged at 300 epochs. Compared with the traditional Yolov5s, the improved Yolov5s showed a lower convergence value and a faster convergence speed (loss: 0.015 vs. 0.031, epochs: 238 vs. 273).

To further verify the effectiveness of different modules in cumulus cell detection, the ablation experiments were conducted, and the Yolov5s networks integrated with different modules were tested on the same dataset. As shown in

Table 1. Results of ablation experiments for the improved Yolov5s.

Model	mAP/(%)	Recall/(%)	FPS/(f/s)
	$\frac{{\displaystyle {\sum }_{\mathit{i}=1}^{\mathit{N}}\mathit{A}{\mathit{P}}_{\mathit{i}}}}{\mathit{N}}$	$\frac{\mathit{T}\mathit{P}}{\mathit{T}\mathit{P}+\mathit{F}\mathit{N}}$	$\frac{1}{\mathit{T}}$
Yolov5s	87.2	81.6	23.3
Yolov5s + Res2Net-COT	88.5 (1.5%)	82.5 (1.1%)	27.5 (18.0%)
Yolov5s + BiFPN	90.6 (3.9%)	83.7 (2.6%)	28.8 (23.6%)
Yolov5s + Inception	87.8 (0.7%)	81.6 (0%)	30.2 (29.6%)
Improved Yolov5s	91.3 (4.7%)	86.4 (5.9%)	26.1 (12.0%)

The number of categories N = 2, AP_i was the average detection accuracy for each category, TP was the number of cumulus cells correctly detected, FN was the number of cumulus cells that were missed, and T was the time required to detect all cumulus cells from an image.

, the experimental results showed that the backbone, neck, and head in the Yolov5s all contributed positively to improving the performance of the network. Adding the Res2Net-COT module to the backbone increased the mAP and recall by 1.5% and 1.1%, respectively. When the BiFPN module was added to the neck of the traditional Yolov5s, the mAP and recall were increased by 3.9% and 2.6%, respectively. The inception module added to the head also contributed to the time cost of the Yolov5s. Although the recall remained the same and mAP was increased by 0.7%, the FPS was significantly increased by 29.6%. After adding the Res2Net-COT module, the BiFPN module, and the inception module to the traditional Yolov5s, the improved Yolov5s achieved the best performance, with the mAP increased by 4.7%, the recall increased by 5.9%, and the FPS increased by 12.0%.

4.4. Success Rate and Discard Rate

The process of cumulus cell removal was considered successful when cumulus cells were completely removed by repeated micropipette aspiration. After 24 hours, oocytes without the presence of the polar body were classified as immature and subsequently discarded. To further quantify the performance of the robotic system, the success rate was calculated as the percentage of the number of oocytes without cumulus cells after cumulus cell removal over the total number of oocytes before cumulus cell removal. The discard rate was calculated as the percentage of the number of discarded oocytes over the total number of oocytes before cumulus cell removal.

In five experiments, cumulus cell removal was conducted on a total of 267 mouse COCs using the proposed robotic system and manual operation (144 vs. 123). As summarized in

Table 2. Experimental results of cumulus cell removal.

Experiment	Manual Success Rate (%)	Robotic Success Rate (%)	Manual Discard Rate (%)	Robotic Discard Rate (%)
1	83.3 (20/24)	96.4 (27/28)	16.7 (4/24)	7.1 (2/28)
2	85.7 (18/21)	100 (25/25)	23.8 (5/21)	4.0 (1/25)
3	88.5 (23/26)	96.7 (29/30)	23.1 (6/26)	0 (0/30)
4	82.6 (19/23)	100 (27/27)	17.4 (4/23)	3.7 (1/27)
5	86.2 (25/29)	97.1 (33/34)	17.2 (5/29)	5.9 (2/34)
Mean ± SD	85.3 ± 2.4	98.0 ± 1.8	19.6 ± 3.5	4.1 ± 2.7

, the success rate of the robotic group was higher than that of the manual group (98.0 ± 1.8% vs. 85.3 ± 2.4%). Additionally, the discard rate of the robotic group was lower compared with manual operation (4.1 ± 2.7% vs. 19.6 ± 3.5%).

4.5. Amplification Rate and Non-Specific Rate

When a cell sample collected from a blastocyst (i.e., day 5/6 embryos) contains cumulus cells, the DNA from the cumulus cells can contaminate the embryonic DNA, reducing the accuracy of the PGT test. The amplification rate and non-specific rate were regarded as two important metrics of the PGT test. The amplification rate was calculated as the percentage of the number of amplified genetic samples containing target bands after gel electrophoresis over the total number of amplified genetic samples. The non-specific rate was calculated as the percentage of the number of amplified genetic samples containing non-specific bands after gel electrophoresis over the total number of amplified genetic samples. The target bands and the non-specific bands were distinguished based on the target band size.

A total of 120 mouse blastocysts were collected after cumulus cell removal and ICSI. Among them, there were 60 blastocysts in the robotic group and 60 blastocysts in the manual group. For each experiment, DNA from five blastocysts was amplified into one target-gene sample by using a single target-gene primer in the PGT procedure. In total, twelve target-gene samples were obtained in either the robotic group or the manual group for gel electrophoresis. The testing results of four target genes are summarized in

Table 3. PGT results of target genes.

Target Genes	Target Band Size (bp)	Amplification Rate (%)		Non-Specific Rate (%)
		Manual	Robotic	Manual	Robotic
E6	644	91.7 (11/12)	100 (12/12)	25.0 (3/12)	16.7 (2/12)
E7	753	75.0 (9/12)	91.7 (11/12)	41.7 (5/12)	0 (0/12)
XBP1	113	83.3 (10/12)	100 (12/12)	16.7 (2/12)	0 (0/12)
β-Actin	96	91.7 (11/12)	100 (12/12)	33.3 (4/12)	8.3 (1/12)
Mean ± SD	—	85.4 ± 8.0	97.9 ± 4.2	29.2 ± 10.8	6.3 ± 6.1

.

The amplification rate of the robotic group was higher than that of the manual group (97.9 ± 4.2% vs. 85.4 ± 8.0%), and the non-specific rate of the robotic group was lower than that of the manual group (6.3 ± 6.1% vs. 29.2 ± 10.8%). The representative banding patterns produced by gel electrophoresis are shown in Figure 11.

5. Discussion

Our results show that the AUC of the EPSANet50 network is more than 0.05 higher than that of other state-of-the-art methods, demonstrating outstanding performance across all the classification thresholds (i.e., from 0 to 1). Compared with the ResNet50 network, the EPSANet50 network can extract multi-scale COC features using the PSA module. Although the DenseNet50 and the SENet50 networks improve the feature extraction by introducing the PSA module and SEWeight module, respectively, they failed to combine these modules effectively. In contrast, the EPSANet50 network integrated the SEWeight module into the PSA module, enriching the extracted COC’s features and paying more attention to key features such as COC’s polar body. Due to the occlusion of the polar body by surrounding cumulus cells, embryologists remove cumulus cells without assessment of oocyte’s maturity, which affects the maturation process and eventually leads to maturation failure. Therefore, it is important that the EPSANet50 network was developed to accurately assess the maturity of oocytes based on the features of the polar body.

Our results suggest that the proposed adaptive controller outperformed LQR and PD controllers in terms of positioning accuracy, overshoot, and settling time. During aspiration, for the PD controller, the k_p and k_d (k_p = 100, k_d = 2600) were obtained by empirical tuning, which did not take the varying mass of the COC into account and thus caused a larger overshoot. Although the LQR controller calculated the k_p and k_d based on the mass of the COC before each aspiration, it ignored the varying mass of the COC during the whole aspiration. In contrast, the proposed adaptive controller updated the mass of the COC in real time during aspiration and calculated k_p and k_d according to Equation (10), compensating for the positioning errors caused by the varying COC mass. Since the mass of the COC varies as the cumulus cell are removed during aspiration, the acceleration of the COC will change if a fixed external force is applied to it, leading to the large positioning errors and the loss of the COC inside the micropipette. Therefore, it is essential that the adaptive controller was designed to compensate for the positioning errors and accurately position the COC at the target position inside the micropipette.

Our results demonstrate that the improved Yolov5s network achieved the best performance of cumulus cell detection. The Res2Net-COT and BiFPN modules contributed to the increase in the mAP and recall, since the Res2Net-COT module acquired the effective semantic information and extracted the fine cumulus cell features, while the BiFPN module was able to address insufficient feature fusion and incomplete feature transmission. The inception module contributed to the increase in the FPS, since its parallel transmission structure effectively improved the inference speed and real-time performance. As the cumulus cells are small, blurry, and varied in size, embryologists cannot accurately calculate the number of the remaining cumulus cells, leading to incomplete cumulus cell removal. Therefore, it is vital that the improved Yolov5s was proposed to quantify both the number and size of cumulus cells, ensuring accurate assessment of the completeness of cumulus cell removal.

We found that the robotic group had higher success and lower discard rates. Compared with the manual group, the accurate assessment of cumulus cell removal was enhanced by the improved Yolov5s in the robotic group. Additionally, the accurate positioning of COCs achieved by the designed adaptive controller increased the success rate of the robotic group, reducing COC loss caused by the large overshoot. The lower discard rate of the robotic group was attributed to the accurate assessment of oocyte’s maturity by the proposed EPSANet50 network before removing cumulus cells from the COCs. Existing methods, such as microfluidic and vortexing devices [13,14], have also been developed for cumulus cell removal. However, in microfluidic operations, it is easy for COCs to become stuck in a narrow microchannel or suffer from incomplete cumulus cell removal in a large microchannel. In vortexing operations, there is a high risk of damaging COCs due to the strong shear force applied.

We also found that the robotic group had higher amplification and lower non-specific rates for the genetic testing. In comparison, cumulus cells were not completely removed in the manual group, and the DNA of cumulus cells reduced the function of Taq-polymerase, which suppressed the generation of target bands in the amplified gene samples. Additionally, more non-specific bands were generated due to the non-specific binding of the cumulus cell DNA to target-gene primers.

Although the robotic system has many advantages, it is not without limitations. The robotic system can process only a single COC at a time. The investigation will be conducted for the manipulation of multiple COCs in future work. Additionally, the current robotic system was tested on mouse COCs. The system should be further validated on human COCs before it can be deployed into the IVF clinics.

As AI-powered tools continue to gain popularity in various fields [39], the future integration of this robotic system into clinical settings appears increasingly promising. The robotic system could significantly enhance the efficiency and accuracy of clinical IVF procedures, particularly for PGT, and reduce the overall human subjectivity, leading to the improvement of the throughput and success rates of IVF treatments. Consequently, the capacity and reputation of the clinics with robotic systems will be further enhanced. Since the robotic manipulator and the camera are designed for the microscopy systems used in IVF clinics, the robotic system can be seamlessly integrated into the clinical setup. Moreover, automated manipulation significantly reduces human labor and minimizes the requirement for manual cumulus cell removal by lab technicians. The technician can complete the cumulus cell removal with just a few mouse clicks via the user interface on a computer without complex manual manipulation.

6. Conclusions

This article reported on a physical robotic system integrated with AI-powered technologies to automate and standardize the removal of cumulus cells from COCs in IVF. An EPSANet50 network was developed to accurately assess the oocyte’s maturity before removing cumulus cells, alleviating the influence on the maturation of the immature oocytes and further reducing the discard rate of oocytes. An adaptive controller was designed to accurately position the COC at the target position inside the micropipette, which took the varying mass of the COC into account and reduced the oocyte loss inside the micropipette. Additionally, an improved Yolov5s network was integrated into the system for quantitatively assessing the cumulus cell number and size, and the completeness of cumulus cell removal, which ensured that the embryonic DNA was not contaminated by the DNA from the cumulus cells in the subsequent PGT procedure. Experiments on mouse COCs revealed a success rate of 98.0 ± 1.8% and a discard rate of 4.1 ± 2.7% by the robotic system. Compared with manual operation, the robotic system achieved a higher amplification rate of 97.9 ± 4.2% and a lower non-specific rate of 6.3 ± 6.1%, indicating the higher performance of the PGT test and its potential to further improve clinical practices in modern IVF.

In future research, we still need to explore new ideas to further automate and standardize our robotic system for cumulus cell removal in IVF. (1) The first idea is creating 3D digital twins for robotic planning prior to the physical control of COCs. This approach involves the 3D modeling of COCs, simulation of cumulus cell removal, and AI-driven path planning and optimization. It allows for the validation of the cumulus cell removal process in a risk-free virtual space, identifying potential risks before any physical manipulation of human COCs prior to clinical trials. (2) The second idea involves developing a cyber–physical approach for the robotic manipulation of COCs. This approach integrates physical robotic manipulators with computational models and vision feedback to remotely and automatically remove cumulus cells. This technology will mitigate the dilemma due to the shortage of experienced embryologists and enable advanced IVF treatments for patients in various locations, including underdeveloped regions.